Artificial Intelligence-Based Traffic Flow Prediction:

A Comprehensive Review
Sayed A. Sayed (  [email protected] )
Cairo University
Yasser Abdel-Hamid
Cairo University
Hesham Ahmed Hefny
Cairo University

Systematic Review

Keywords: ITS, AI, Traffic Prediction, Traffic Congestion, Machine Learning, Deep Learning

Posted Date: July 28th, 2022

DOI: https://doi.org/10.21203/rs.3.rs-1885747/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Page 1/50
Abstract
The expansion of the Internet of Things has resulted in new creative solutions, such as smart cities, that
have made our lives more productive, convenient, and intelligent. The core of smart cities is the Intelligent
Transportation System (ITS) which has been integrated into several smart city applications that improve
transportation and mobility. ITS aims to resolve traffic issues, especially traffic congestion. Recently, new
traffic flow prediction models and frameworks have been rapidly developed in tandem with the
introduction of Artificial Intelligence (AI) approaches to improve the accuracy of traffic flow prediction.
Traffic forecasting is a crucial duty in the transportation industry. It can significantly affect the design of
road constructions and projects in addition to its importance for route planning and traffic rules.
Furthermore, traffic congestion is a critical issue in urban areas and overcrowded cities. Therefore, it must
be accurately evaluated and forecasted. Hence, a reliable and efficient method for predicting traffic is
essential. The main objectives of this study are: First, present a comprehensive review of the most
popular machine learning and deep learning techniques applied in traffic prediction. Second, identifying
inherent obstacles to applying machine learning and deep learning in the domain of traffic prediction.

1. Introduction
In recent decades, the demand for the development of ITS-based solutions for precise traffic prediction
and mobility management has increased as cities have gotten increasingly crowded and congested [1].
ITS is an advanced technology for delivering transportation by utilizing advanced data communication
technologies through the integration of communications, computers, information, and other technologies
and applying them to the transportation industry. This process aims to create an integrated system of
people, roads, and vehicles [2]. ITS can construct a comprehensive, real-time, accurate, and effective
transportation management system [3]. Furthermore, it has the potential to significantly reduce hazards,
high accident rates, traffic congestion, carbon emissions, and air pollution, while also improving safety
and dependability, travel speeds, traffic flow, and passenger satisfaction [4].

Precise traffic flow prediction is essential to the ITS as it can help traffic stakeholders (Individual
passengers, traffic administrators, policymakers, and road users), shown in Fig. 1, utilize transport
networks more safely and intelligently [5,6]. The efficacy of these systems depends on the quality of
traffic data; only then an ITS will be successful. According to the World Health Organization's (WHO) 2018
report on the universal status of road safety, road traffic deaths continue to rise, with 1.35 million deaths
recorded in 2016, making the study of traffic forecasting a valuable method for reducing congestion and
ensuring safer, more cost-effective travel [7,8]. The benefits of traffic flow forecasting are illustrated in Fig.
2.

Historically, traffic flow forecasting was dependent on parametric models such as time series analysis
derived from historical data. In time series, a collection of observed readings x is recorded at a specific
time t. The objective is to recognize temporal patterns in past traffic data and use these results for
forecasting. Another model for mobile stochastic problems capable of resolving regression concerns and

Page 2/50
minimizing variance to achieve optimal results was the Kalman Filtering method for time-series analysis
[9]. Also, the Auto-Regressive Integrated Moving Average (ARIMA) model is a well-known and standard
framework for predicting short-term traffic flow [10]. Numerous modifications to the ARIMA model were
implemented, and the results ensured an enhanced performance [11-13].

Because traffic flow is stochastic and non-linear, non-parametric models such as Random Forest (RF)
Algorithm, Bayesian Algorithm (BA) approach, K-Nearest Neighbor (KNN), Principal Component Analysis
(PCA), and Support Vector Algorithms [9] have recently been employed in traffic flow prediction. In
addition, neural networks became popularly employed for predicting traffic flow [15]. In the era of big
data, a shallow Back-Propagation Neural Network (BPNN) [16] showed promising results, unfortunately, it
failed. Thus, deep learning emerged, which employs several layers to progressively extract more complex
properties from raw input. Convolutional Neural Network (CNN) [17], Recurrent Neural Network (RNN) [18],
Long Short-Term Memory (LSTM) [19], Restricted Boltzmann Machines (RBM) [20], Deep Belief Network
(DBN) [21], and Stacked Auto Encoder (SAE) [22] are some examples of deep learning architectures.

The primary goals of this research are to conduct a comprehensive survey of the key machine learning
and deep learning techniques used in forecasting traffic flow in addition to identifying the obstacles and
future directions for machine learning and deep learning in this field.

The rest of the paper is organized as follows: Section 2 gives a theoretical background about traffic
prediction problem, machine learning, and deep learning. Section 3 outlines the survey methodology.
Section 4 presents a literature review of machine learning and deep learning approaches employed in
traffic flow prediction. Section 5 covers the existing challenges in the topic of this survey. Finally, Section
6 concludes the paper.

2. Background
ITS provide a punch of high-resolution traffic data to be used in data-driven based traffic flow prediction
techniques [23]. From this perspective, traffic flow prediction can be considered as a time series problem
in which the flow count at a future time is estimated based on data received from one or more
observation points during prior periods. Traffic flow forecasting is a major component of traffic modeling,
operation, and management. Accurately predicting traffic flows in real-time can give information and
recommendations for road travelers to enhance their travel choices and decrease expenses, in addition to
supplying authorities with enhanced traffic control tactics to alleviate congestion. Machine learning and
deep learning as depicted in Fig. 3, are considered as subsets of AI that have witnessed exponential
expansion over years [24]. These approaches have been deemed successful in predicting traffic flow.

2.1 Machine Learning

Page 3/50
Machine Learning (ML) techniques are considered statistical models that are utilized to make
classifications and predictions based on the data provided [24]. ML is an area of AI that focuses on the
development of prediction algorithms depending on the fair discovery of patterns within huge datasets
and without being designed specifically for a particular job [25]. ML models are classified into three
categories according to the learning techniques they employ: supervised learning, unsupervised learning,
and reinforced learning (RL). In addition, ML algorithms might be further subdivided into several
subgroups depending on distinct learning approaches, as shown in Fig. 4 [26].

2.1.1 Supervised Learning

In the tasks that depend on supervised learning, a labeled dataset known as feature vectors and their
corresponding predicted output labels are supplied to the model. The objective of these models is to
create an inference function that maps feature vectors into output labels. When the ML model training is
complete, it can make predictions based on new data. Continuous or discrete predictions can be
generated using supervised learning algorithms [24]. Support Vector Machine (SMV), KNN, Logistic
Regression, Linear Regression, Decision Trees (DT), Random Forests (RF), and Naive Bayes are examples
of supervised learning approaches [25].

A. Support Vector Machine

SVM is a supervised learning methodology based on the classification approach. It can be considered a
non-probabilistic linear classifier. SVM is regarded as the state-of-the-art machine learning algorithm.
Margin calculation is the core concept underlying SVM. In such an approach, each item of data is
represented in n-dimensional space as a point, where n is the features count and each feature represents
the value of the coordinate. As depicted in Fig. 5, the objective of this strategy is to examine the
vectorized data as well as create a hyperplane that distinguishes between the two classes [27]. Various
margins are then drawn between several classes, and a hyperplane is built that minimizes the mean-
squared error and maximizes the margin-to-class distance [28].

Once an optimal separating hyperplane is identified in the case of linearly separable data, points of data
that sit on its boundary are called support vector points, and the solution is introduced as a linear
combination of these points alone, as depicted in Fig. 6. The other data values are disregarded [29].
Therefore, the SVM model's complexity is independent of the features count found in the training data.
So, SVMs are ideally suitable for learning missions involving a large number of features relative to the
number of training cases.

Page 4/50
Despite the greatest margin that enables the SVM to choose through numerous nominee hyperplanes,
SVM may be unable to locate any hyperplane that can separate hyperplanes at all due to the
misclassified instances contained in the data. One proposed solution to this problem is to utilize a soft
margin that allows certain training cases misclassifications [30]. SVMs are binary classifiers, so in the
case of multi-class problems, the problem needs to be reduced to a series of several binary classification
problems. Categorical data represents another challenge, however, with adequate rescaling, decent results
can be obtained [29].

B. K-Nearest Neighbors

KNN is considered a non-parametric classification technique that makes no assumptions about the basic
dataset and is known for its efficiency and simplicity. In KNN, a labeled training dataset is used to predict
the class of unlabeled data [31]. KNN is typically employed as a classifier to classify data based on the
nearest or most nearby training samples in a specific location. KNN is utilized in datasets where data
may be divided into distinct clusters to determine the new input’s class. KNN is more significant in case
there is no prior knowledge of the data used in the study [31].

KNN typically employs K variable values between 0 and 1 to calculate the number of training data points
with the closest distance. KNN employs numerous distance functions, including Manhattan distance,
Euclidean distance, Minkowski distance, and Hamming distance. The Euclidean distance is employed to
calculate its nearest neighbors in the case of continuous data, but for categorical data, the Hamming
distance function is utilized [32].

The most challenging aspect of the KNN algorithm is choosing the K value, as it affects the algorithm's

performance and precision. Small K values generate noise in class label prediction, while large K values
may lead to excess fitting likelihood. In addition, it increases the computation time and affects the
execution speed. The K value is calculated according to (1):

K = n^(1/2) (1)

Where n is the size of the dataset.

Cross-validation will be applied to training data with varied K values to maximize the test results. The
optimal value for test results will be decided based on the optimal precision [32].

Page 5/50
The KNN technique has the following benefits: it is a straightforward technique that is simple to apply. It
is a very adaptable classification technique that is ideal for multimodal classes.

On the other hand, using the KNN algorithm to classify unknown data is quite costly. It needs to calculate
the distance between the k-nearest neighbors. As the size of the training set increases, algorithm
computations get increasingly intensive. Noisy or irrelevant characteristics will decrease accuracy.
Moreover, KNN does no generalization on the training data and retains them all. Consequently, greater
dimensional data will reduce the precision of areas. It computes the distance between k neighbors, so
KNN is a lazy learner [33].

C. Logistic Regression

Logistic regression is a supervised learning approach used to differentiate between two or more groups
[27]. It provides, in terms of 0 and 1, the likelihood that an event will occur based on the values of the
input variables (i.e., it gives the binomial outcome). For instance, predicting whether or not an e-mail is
categorized as spam is a binomial result of Logistic Regression. In addition, Logistic Regression can
produce multinomial outcomes, such as predicting the preferred cuisine (Chinese, Italian, Mexican, etc.).
In addition, Logistic Regression can produce ordinal results, such as rating a product from 1 to 5, etc.
Therefore, Logistic Regression is concerned with categorical target variable prediction [33]. Logistic
Regression provides several benefits, including ease of implementation, computational efficiency, training
efficiency, and regularization simplicity. In Logistic Regression, input features do not require scaling. In
addition, Logistic Regression is immune to data noise and multi-collinearity. Logistic Regression, on the
other hand, is unsuitable for non-linear problems since its decision surface is linear, sensitive to
overfitting, and all independent variables must be recognized for it to work successfully [33].

D. Linear Regression

Regression is an example of a supervised learning technique in which the value of the output variable is
decided by the values of the input variable and the utilized labeled datasets. Regression can be used to
model and predict continuous variables. In linear regression, an attempt is made to fit a straight
hyperplane to the data set if the relationship between the variables of a dataset is linear [33]. Linear
Regression is calculated according to (2) [32]:

Page 6/50
 F (x) = mx + b + e (2)

Where x is the independent variable, F(x) is the dependent variable, m is the slope of the line, b is the y-
intercept, and e is the error term.

The best prediction accuracy may be achieved using the Linear Regression algorithm if the following
steps are followed to prepare the training data [32]:

Assume that the dependent and independent variables are linear, i.e., apply any of the available data
transformation techniques to make the data linear.
Remove noisy data and outliers using a technique for cleaning data.
To minimize overfitting, do pair-wise correlation and exclude the most linked variables.
Apply Gaussian distribution to the training data to generate more accurate predictions.
Rescale inputs to improve the reliability of the prediction.

From the above discussion, it is clear that the Linear Regression algorithm is straightforward to
comprehend. In addition, the ideal linear relationship between dependent and independent variables is
demonstrated. In contrast, Linear Regression can only predict the numeric output. It is inappropriate for
nonlinear data and highly sensitive to outliers. Also, data must be independent [32].

E. Decision Trees

Classifier-generating systems are one of the most popular strategies in data mining [34]. In data mining,
classification algorithms are capable of processing vast quantities of data. It can be used to create
assumptions about categorical class names, categorize information based on training sets and class
labels, and classify newly accessible data [35].

DTs are one of the powerful approaches utilized in numerous domains, including ML, image processing,
and pattern recognition [36]. DT is a model that sequentially as well as cohesively combines a set of
basic tests in which a numerical characteristic is compared with a threshold value [37]. In addition, DT is
a common classification model in Data Mining [38]. Every tree is composed of nodes and branches. Each
node represents an attribute inside a group to be categorized, and each branch provides a possible value
for the node [39]. Fig. 7 illustrates the structure of DT.

DT algorithm is a supervised learning algorithm. It tries to build a training model that may be used to
predict the class or value of target variables by employing learning decision rules learned from the
Page 7/50
training data [41].

The advantages and disadvantages of using the DT algorithm to solve regression and classification
problems [42 - 44] are outlined in Table 1.

Table 1 DT Benefits and Drawbacks

Benefits Drawbacks
1- Easy to understand. 1- The ideal decision-making system can be
2- Rapidly transformed into a set of production principles. thwarted, resulting in wrong decisions.
3- Capable of classifying categorical and numerical 2- The decision tree has multiple levels, which
outcomes, but only capable of generating categorical makes it confusing.
attributes. 3- The complexity of the decision tree's
4- There are no a priori hypotheses on the validity of the calculations may increase as training samples
outcomes. increase.

F. Random Forest

RF is an ensemble classifier since it employs a large number of DTs to compensate for the shortcomings
of a single DT [45-49]. The 'vote' of all trees is utilized to determine the final class for each unknown. This
eliminates the possibility that a single tree may not be ideal. Therefore, adding numerous trees should
result in a global optimum [50]. For the formation of each tree in the "forest", the bootstrap approach is
used for resampling. In addition, on each node split, a subset of features is randomly selected and the
split variable selection occurs over this subset. The projected value for classification is the majority vote,
and the average, for regressions [51-54]. On RF models, there are two parameters for tuning: mtry, which
is the number of features that are randomly picked to consider in each split; and ntree, which is the trees
count in the model. The mtry parameter has a tradeoff: large values increase the correlation among trees
but improve the accuracy of each tree [51]. The unused elements are called the Out of Bag (OOB)
samples, which can be employed for validation in this case, each tree predicts over its OOB samples and
the final result is an average over the outcomes of the trees [55].

There are two options for estimating the relevance of each variable and ranking them accordingly. The
initial choice is to utilize the OOB samples. In this option, the accuracy is calculated over the set of each
tree and its corresponding OOB samples, a variable is randomly permuted among samples, and the
accuracy is recalculated on the new set. Applying this to the set of all trees and average for each variable
yields a metric for comparing relevance. This metric for comparison is known as the Permutation
Importance Index (PIM) or Variable Importance Measure (VIM). The alternative is to calculate the split
improvement for each tree and node using a measure (e.g., the Gini Index) and use these values to
compare the significance of the variables [55].

RFs offer high flexibility and prediction rates. It also does not overfit the data when the number of trees is
considered. Alternatively, a graphical representation is not feasible as in DTs [55].

Page 8/50
G. Naïve Bayes

The Naïve Bayes technique, also known as the Bayes of Idiots, Bayes of Freedom, or basic Bayes, is a
fundamental probability-band classifier. Provided the class variable, it is supposed that the existence or
absence of a particular class feature has no significance on the existence or absence of any other class
feature [56].

The Naïve Bayes technique is straightforward to implement since it does not require complex recursive
parameter estimation systems. Consequently, a naive Bayes classifier can be useful for enormous
datasets. Also, it requires minimal training data to assess the restrictions. As independent variables rather
than the whole matrix of covariance are assumed, only the variances of the variables within each class
must be estimated [56].

2.1.2 Unsupervised Learning

In unsupervised learning, there is no output label information contained in the dataset. The purpose of
these models is to infer the link between data and/or to uncover hidden variables [25]. These strategies
are mostly used to reduce the size of a dataset by extracting key features. Reducing the number of
features helps prevent problems such as high computational cost and multi-collinearity [57]. Fig. 8
depicts unsupervised learning, in which the machine guesses the result according to past experiences
and learns from information previously provided to anticipate the real-valued outcome. Examples of
unsupervised learning-based methods are K-Means Clustering, Principal Component Analysis (PCA), and
Latent Dirichlet Allocation (LDA) [25].

A. K-Means Clustering

K-means clustering is one of the unsupervised learning methods that automatically produces groups or
clusters. Data with comparable properties are put into the same cluster. K-means is the name of the
method as it forms K different groups [28]. The purpose of the K-means clustering is twofold: (1) to
provide K-centroids, one for each cluster, and (2) to minimize the square error function. The mean value is
placed in the middle of the cluster [27].

Page 9/50
The k-means clustering technique has many advantages. First, it is computationally more effective than
hierarchical clustering for enormous variables. Second, it yields tighter clusters than hierarchical
clustering with global clusters and small k. Finally, ease in implementation and comprehension of the
clustering results. The order of complexity of the algorithm is O(K*n*d), so it is computationally efficient
[33].

On the other hand, the K value is not known and its prediction is complex. Degrades in performance occur
when clusters are global and when different beginning partitions result in distinct final clusters. Also,
when there is a difference in the size and density of the clusters in the input data, the performance
decreases. In addition, the joint distribution of characteristics inside each cluster is spherical (spherical
assumption) and cannot be achieved as the correlation between features break it and put extra weights
on connected features. K-Means clustering can be susceptible to outliers. Also, it is sensitive to local ideal
and initial points, and a unique solution for a specific K value does not exist - so K means needs to be run
for a K value lots of times (20-100times) and then pick the results with the lowest J [33].

B. Principal Component Analysis

PCA is an unsupervised ML approach that reduces the dimension of the data. Therefore, the
computations are more efficient and quicker [27]. The two-dimensional data in PCA is turned into one-
dimensional data by transforming the collection of variables into new ones called principal components
(PC) which are orthogonal. The data set of the PCA algorithm must be scaled because the results are
sensitive to the relative scaling [28].

To explain the PCA mechanism, let’s use an example of 2D data. When the 2D data is plotted on a graph,
it takes up two axes. Applying PCA to this data will turn it into 1D [27], as illustrated in Fig. 9.

C. Latent Dirichlet Allocation

LDA is a statistics-based data mining technique that differentiates between classes of objects in N-
dimensional feature space by computing a sequence of k ≤ N - 1 linear discriminant whose values can be
used to describe the classes [59]. LDA and PCA are similar [60] in that they describe the "most important"
variations in the data and select directions that maximize feature variance. LDA differs from PCA in that
LDA makes use of the class labels: it selects directions that can best differentiate the class means
relative to the sum of the class variances along that direction. It maximizes the ratio of between-class
scatters to within-class scatter. Intuitively, it detects lower-dimensional descriptions of the data which
push the class members together and pull members of different classes out [61]. The k linear

Page 10/50
discriminants that correspond to the eigenvectors are arranged by eigenvalue. The discriminants can be
used to group new objects or for dimension reduction [61].

To ensure the discriminant's optimality, the LDA's design makes the following two assumptions: 1) the
linear combination of any characteristics are normally distributed, and 2) the classes have equal
covariance matrices. Despite the danger of inferior outcomes, LDA has been utilized routinely for
dimension reduction and classification when these assumptions are broken [61].

2.1.3 Reinforced Learning

Unlike supervised and unsupervised learning, RL is a goal-oriented learning approach. Learning occurs via
reacting to the surrounding environments and detecting status changes. RL is strongly tied to an agent
(controller) responsible for the learning process to attain a goal. In particular, the agent takes actions
(control signals) and consequently, the status of the environment is changed and rewards, which are
special numerical values, are returned either positive or negative. The agent aims to maximize the
rewards obtained over time. A task is a full specification of an environment, which determines how the
reward is generated [62]. Examples of RL-based techniques are Q-Learning Algorithm and Monte-Carlo
Tree Search (MCTS).

A. Q-Learning

Q-learning [63] is a straightforward way that enables agents to learn how to act optimally in controlled
Markovian domains. It represents an incremental approach to dynamic programming which imposes low
processing demands. It works by boosting successively its ratings of the quality of specific acts at certain
states. It can also be considered an asynchronous Dynamic Programming (DP) approach. It provides
agents with the possibility of learning to act optimally in Markovian domains by experiencing the
consequences of actions, without requiring them to generate maps of the domains [64].

Q-learning is applied in information theory, and related investigations are underway. Recently, Q-learning
and information theory have been applied to various disciplines such as natural language processing,
anomaly detection, pattern recognition, and image classification [65], [66], [67], [68]. In addition, a
framework has been established to provide a satisfying response based on the user’s speech using RL in
a voice interaction system [69], and a high-resolution prediction system for local rainfall based on DL has
been developed [70].

The advantage of the ant Q-learning approach is that it can identify the value of the reward for a specific
activity in a multi-agent environment successfully due to the corporation between agents. The drawback

Page 11/50
of ant Q-learning is that its result can be stuck at a local minimum when agents take just the shortest
path [71].

B. Monte Carlo Tree Search

MCTS is a powerful technique for handling sequential decision problems. The plan relies on a smart tree
search that balances exploration and exploitation. Random sampling is employed in MCTS in the form of
simulations to save statistics of activities and make more knowledgeable selections in each future
iteration [72]. MCTS is a decision-making technique that is utilized in scanning huge combinatorial
spaces represented by trees. In such trees, nodes represent states, also referred to as configurations of
the problem, whereas edges denote transitions (actions) between states [72].

Formally, MCTS is directly applied to issues that can be described by a Markov Decision Process (MDP).
Certain modifications of MCTS make it possible to be applied to Partially Observable Markov Decision
Processes (POMDP) [73]. More recently, MCTS paired with deep RL are considered the backbone of
AlphaGo developed by Google DeepMind which is documented in [74].

The basic MCTS procedure is conceptually so simple [75], as depicted in Fig. 10. A tree is created in an
incremental and unbalanced method. In each iteration, a tree policy is utilized to get the most urgent node
of the current tree.

The tree policy aims to balance the considerations of exploration and exploitation. A simulation is then
run from the specified node and the search tree result is accordingly updated. This involves the insertion
of a child node that matches the action taken from the selected node and an update of the statistics of
its ancestors. Based on some default policy, moves are being conducted during this simulation which in
the simplest scenario aims to make uniform random moves. A notable advantage of MCTS is there is no
need for the values of the intermediate states to be evaluated, which extremely minimizes the amount of
domain knowledge required [75].

2.2 Deep Learning

About a decade ago, Deep Learning (DL) emerged as an effective ML technique and achieved good
performance in several application fields. The core idea of DL approaches is to learn complicated
characteristics extracted from data with low external contribution using Deep Neural Networks (DNN)
[77]. These algorithms don’t require to be manually provided created features; they automatically learn
additional complicated features [78].
Page 12/50
DL is an AI paradigm that has gained major interest from the academic community and demonstrated
higher potential over conventional methods [79]. DL is a more efficient, monitored, time-consuming, and
cost-effective technique than the ML technique. Not only it is a specific approach to knowledge, but also
it adapts to various methodologies and topographies that could be beneficial to a wide range of
complicated problems. The approach learns the illustrative and differential properties in a relatively
varied method [80, 81]. Fig. 11 demonstrates the procedure of ML and DL.

To generate high-level abstractions with many nonlinear transformations, DL is based on a collection of

ML techniques used to model data. The artificial neural network (ANN) system runs on a DL technology
[82, 83]. These networks include many layers for collecting high-level characteristics and for eliminating
problematic data, so the performance of DL algorithms is higher than ML algorithms [84].

ML approaches have brought a huge impact on our daily life such as efficient web search, self-driving
vehicles, computer vision, and optical character recognition. Also, by implementing ML approaches, the
human-level AI has been improved as well [85-87]. Nevertheless, the performance of classic ML
algorithms is far from ideal when it comes to human information processing mechanisms (e.g., voice and
vision). The DL algorithms concept was formed in the late 20th century inspired by deep hierarchical
structures of human speech perception and production systems. Fig. 12 displays a timeline showing the
evolution of deep models along with the classic model [26]. DL has many architectures. Examples of
such architectures are CNN, RNN, LSTM, and Recurrent CNN (RCNN).

A. Convolutional Neural Network

CNNs are a subtype of ANNs and are frequently utilized in face recognition, text analysis, human organ
localization, and biological image recognition [88]. CNN structure was first introduced in 1988 by
Fukushima [89]. It was not widely employed, however, due to restrictions of computation gear for training
the network. In the 1990s, LeCun et al. [90] adapted a gradient-based learning algorithm to CNNs and
provided successful results for the handwritten digit classification problem. After that, researchers
progressively enhanced CNNs and reported state-of-the-art results in different recognition tasks.

A CNN architecture includes three components: input layer, hidden layer, and output layer. The
intermediate levels of any feedforward network are known as hidden layers, and their number varies
based on the network architecture type. Convolutions are executed in the hidden layers, which include dot
products of the convolution kernel with the input matrix. Each convolutional layer generates feature maps
to be used as input by the subsequent layers [91], as shown in Fig. 13.

In general, CNNs consist of two major components: Feature extractors and a classifier, as shown in Fig.
14. In the feature extraction layers, each layer of the network takes as its input the output of its immediate
previous layer and transmits its output to be the input to the next layer. The CNN design involves a
combination of three types of layers: Convolution, max-pooling, and classification. In the low and middle-

Page 13/50
level of the network, there are two types of layers: Convolutional layers and max-pooling layers.
Convolutions are the even-numbered layers whereas the odd-numbered layers are for max-pooling
operations. The output nodes of the convolution and max-pooling layers are then arranged into a 2D
plane named feature mapping. Usually, the plane of each layer is produced by the combination of one or
more planes of the previous levels. The nodes of a plane are connected to a small section of each
connected plane of the previous layer. Each node of the convolution layer extracts the features from the
inputs by convolution operations on the input nodes. As the features propagate to the highest level, the
dimensions of the features are lowered based on the kernel size of the convolutional and max-pooling
processes correspondingly.

For ensuring classification accuracy, the number of feature maps is increased for expressing better
features of the input. The output of the last CNN layer is used as the input to a fully connected network
called the categorization layer. In the classification layer, the extracted features are being used as inputs
concerning the size of the weight matrix of the final neural network. At the topmost classification layer,
and using a soft-max layer, the score of the respective class is calculated. According to the highest score,
the classifier produces output for the corresponding classifications [92].

CNNs have various advantages including being more like the human visual processing system, having a
highly optimized structure for processing 2D and 3D images, and being effective in learning and
extracting abstractions of 2D information. The max-pooling layer of CNNs is successful, particularly at
absorbing shape variations. Furthermore, CNNs contain much fewer parameters than a fully connected
network of the same size as it is constructed of sparse connections with coupled weights. In addition,
CNNs are trained with the gradient-based learning technique that suffers less from the diminishing
gradient problem. Given that the gradient-based technique trains the full network to reduce an error
criterion

directly, CNNs can generate highly optimized weights [92].

B. Recurrent Neural Network

Developed in the 1980s, RNN is one of the most widely used DL models [93]. These kinds of networks
have a memory that stores the information they have seen so far, and have various types. Moreover,
RNNs are powerful models for time series analysis, and they use the prior output to predict the next
output. In this situation, the networks themselves contain repeating loops in the hidden layers, which
allow the storing of previous input information for a while, so that the system can predict future outputs.
The output of the hidden layer is retransmitted t times to the hidden layer. The output of a recursive layer
is only sent to the next layer when the number of iterations is completed. In such a circumstance, the
Page 14/50
output is more global, and the preceding knowledge is maintained for longer. Finally, the errors are
returned backward to update the weights [94]. RNN is employed mostly in the fields of speech processing
and Nature Language Processing (NLP) settings [95,96].

Unlike CNN, RNN employs sequential data in the network. As the embedded structure in the data
sequence gives useful information, this property is fundamental to a range of various applications such
as NLP. Thus, RNN can be considered as a unit of short-term memory, where x is the input layer, y is the
output layer, and s represents the state (hidden) layer [97]. For a specific sequence of input, a typical
unfolded RNN diagram is presented in Fig. 15. In addition, a deep RNN was introduced to minimize the
learning difficulty in the deep network and brings the benefits of a deeper RNN depending on three
different deep RNN techniques, namely "Hidden-to-Hidden", "Hidden-to-Output", and "Input-to-Hidden"
introduced by Pascanu et al [98].

One of the main challenges with RNN is its sensitivity to the expanding gradient and vanishing problems
[99]. More specifically, the reduplications of many large or small derivatives during the training phase
may cause exponentially explode or decay of the gradients. With the introduction of new inputs, the
network stops thinking about the original ones; hence, its sensitivity decays over time [97].

C. Long Short-Term Memory

LSTM is a special case of RNN as it has internal memory and multiplicative gates. The diversity of LSTM
cell layouts has been described in 1997 when the first LSTM was launched [100]. LSTM contributed to
the development of well-known services like Siri, Cortana, Alexa, Google Translate, and Google voice
assistant [101]. LSTM is a module in an RNN network that addresses missing gradient problems.
Generally, RNN employs the LSTM network to avoid propagation errors. This allows the RNN to learn
across multiple time steps. LSTM includes cells that keep information outside of a recurrent network.
Like the memory in a computer, the cell is deciding when the data has to be stored, written, read, or
erased using the gate [102]. A simple RNN cell depicted in Fig. 16(a) was enhanced by adding a memory
block which is controlled by input and output multiplicative gates. Fig. 16b shows the LSTM architecture
of the jth cell cj. The main component of a memory block is the self-connected linear unit sc termed
constant error carousel (CEC) which protects LSTM from the drawbacks of regular RNN. An input gate
and output gates consist of corresponding weight matrices and activation functions [101].

Generally, it can be concluded that the LSTM cell comprises one input layer, one output layer, and one
self-connected hidden layer. The hidden layer may contain 'conventional' units that can be fed into the
next LSTM cells. However, a conventional LSTM cell also met some limits due to a linear form of sc. It
was specified that its steady expansion may induce saturation of the function hand converted into an

Page 15/50
ordinary unit. Therefore, an additional forget gate layer was inserted [103], as illustrated in Fig. 16(b),
which permits undesirable information to be wiped and forgotten.

Bidirectional LSTM, Hierarchical LSTM, Convolutional LSTM, Grid LSTM, LSTM Autoencoder, and Cross-
modal LSTM are the most advanced network topologies that use the LSTM gating mechanism [104].

Bidirectional LSTM type networks send and receive the state vector in both directions. As a result, bi-
directional time dependencies are taken into account. As a result of reverse state propagation, future
expected correlations can be included in the network's generated outputs. Hence, more time dependencies
can be detected, extracted, and resolved using bidirectional LSTM networks more precisely than
unidirectional LSTM networks. LSTM networks can encapsulate geographically and temporally dispersed
information and harmonize partial data using a flexible connection mechanism for the propagation of the
cell state vector [105]. Based on the data gaps discovered, this filter method redefines the connections
between cells. Fig. 17 depicts the architecture of Bidirectional LSTM.

Hierarchical LSTM networks resolve multidimensional problems by splitting the overall problem into
sub-problems and hierarchically structuring them. This is achieved by adjusting weights inside the
network which obtains the power to produce a specific degree of attention.

Using a weighting-based attention mechanism that handles and filters input sequences, hierarchical
LSTM networks could be utilized to predict long-term dependencies [106]. Convolution LSTM can be used
to filter and reduce input information obtained over a longer time period using convolution operations
built into LSTM networks or directly into the LSTM cell structure. Convolution methods that are directly
incorporated into the cell can also be used to extend the usual LSTM cell. Correlations are extracted by
convolving current input sequences, recurrent output sequences, and weight matrices. The newly created
features are received as new inputs by the network gates [107]. Fig. 18 depicts this strategy.

Moreover, convolutional LSTM networks are considered ideal for expressing a wide range of quantities,
including spatially and temporally distributed relations. Nevertheless, as a reduced feature representation,
various values can be collectively forecasted alone. Layers' deconvolving must predict different output
quantities based on their original units rather than as features [104]. An autoencoder structure is
commonly used to realize information deconvolution and convolving. A layered LSTM autoencoder
handles the challenge of high dimensional input data and the forecasting of high dimensional parameter
spaces in [108]. In [109] a method for directly integrating an autoencoder into the LSTM cell structure was
proposed. This multimodal prediction approach was proposed by extending LSTM. To compress input
data as well as cell states, encoders and decoders were integrated directly into the LSTM cell structure.
This optimization maximizes information flow in the cell and leads to an enhanced cell state update
mechanism for both short-term and long-term dependencies.

Grid LSTM is an LSTM cell with a matrix structure [110]. The Grid LSTM has connections for the
input sequences' spatial and temporal dimensions. As a result, connections in various dimensions within
cells extend the normal information flow. As a result, the Grid LSTM is appropriate for parallel prediction
Page 16/50
of a wide range of output quantities that can be either linearly independent or nonlinearly dependent. Fig.
19 compares a two-dimensional Grid LSTM network to a standard stacked LSTM network [110].

Cross-modal LSTM is a modern method for predicting various quantities collaboratively. It combines a
number of regular LSTMs that were previously used to separately simulate the individual quantities. The
LSTM flows interact via recurrent connections to handle the quantity dependencies. In other streams, the
outputs of defined layers are used as extra inputs for previous and subsequent layers. As a result, a cross-
modal prediction can be identified. Fig. 20 depicts cross-modal LSTM [111].

D. Recurrent Convolution Neural Network

In recent years, a new class of CNNs, RCNN, inspired by rich recurrent connections in the visual systems
of animals, was introduced. The main component of RCNN is the recurrent convolutional layer (RCL),
which integrates recurrent connections across neurons in the normal convolutional layer. With the
increasing number of recurrent computations, the receptive fields (RFs) of neurons in RCL expand
unboundedly, which is incongruous with biological realities [112]. The traditional RCNN model was
proposed in [113, 114]. The RCNN architecture is presented in Fig. 21, in which both feed-forward and
recurrent connections have local connectivity and shared weights across distinct locations. This
architecture is quite close to the recurrent multilayer perceptron (RMLP) which is generally used for
dynamic control [115, 116] (Fig. 21, middle). The main difference is that the full connections in RMLP are
replaced by shared local connections, similar to the difference between MLP [117] and CNN.

RCNN integrates a stack of RCLs, optionally interleaved with max-pooling layers, as seen in Fig. 22. Here,
layer 1 is the traditional feed-forward convolutional layer without recurrent connections, followed by max
pooling. Furthermore, four RCLs are employed with a max-pooling layer in the middle. There are only feed-
forward connections among nearby RCLs. Both pooling operations have stride 2 and size 3. The output
of the fourth RCL follows a global max-pooling layer, which yields the maximum across every feature
map, providing a feature vector describing the image. Finally, a softmax layer is utilized to categorize the
feature vectors into C categories. [113].

RCNN has various advantages from the computational perspective. First, the recurrent connections
in RCNN allow every unit to include context information in an arbitrarily broad region in the current layer.
Second, the recurrent connections improve the depth of the network and at the same time keep the
number of changeable parameters constant by sharing weight. This is compatible with the tendency of
the current CNN architecture. Third, unfolded RCNN is a CNN with numerous paths from the input layer to
the output layer, which facilitate learning. On one hand, the existence of longer paths makes the model
capable of learning more complicated features. On the other hand, the existence of shorter paths may
improve gradient backpropagation during training [113].

3. Survey Methodology
Page 17/50
The articles reviewed in this paper have been published in high-quality conferences and journals of IEEE,
Elsevier, Springer, and IOP publishing. Machine learning, deep learning, traffic flow prediction, traffic flow
forecasting, traffic speed prediction, short-term traffic prediction, short-term traffic forecasting, and ITS are
some of the search terms used to find these articles. The articles examined in this survey are directly
relevant to the application of ML and DL approaches in traffic flow prediction. Both empirical and review
literature on the abovementioned subjects were considered for this work.

3.1 Survey Organization

This survey compares various forecasting techniques for traffic flow. It follows a dual structure with ML
techniques used for traffic flow prediction and DL techniques utilized for traffic flow prediction. This study
provides a detailed discussion of the approaches and algorithms which are utilized for predictions,
performance measurements, and tools used for these procedures.

The prediction of traffic flow has become one of the primary tasks in the ITS field [118]. Statistical
methods, AI, and data mining techniques have been widely employed recently to evaluate road traffic
data and anticipate future traffic indicators [119]. Previous findings demonstrated that no single
technology is capable of evaluating enormous datasets only by itself. Therefore, according to the data
structure and its volume, the proper technology has to be applied to extract the best insight from the
collected data [120].

3.1.1 ML techniques for Traffic Flow Prediction

In [121], the authors developed an ML-based traffic flow prediction paradigm employing a regression
model implemented by several libraries including Pandas, Numpy, OS, Matplotlib, Keras, Sklearn, and
Tensorflow. Traffic prediction in this study involves the prediction of next year’s traffic data based on
previous years' traffic data which eventually offers the accuracy and mean square error. The traffic
information was predicated on a basis of 1-hour time gap. Data in this study was acquired from the
Kaggle dataset. Two datasets were obtained, in which one is the 2015’s traffic data which contains the
date, time, number of cars, and number of junctions. The other one is the 2017’s traffic data with identical
specifications to compare easily without any confusion. This study needs to investigate more aspects
that affect traffic flow prediction and employ other prediction approaches like deep learning and big data.

In [122], the authors aimed to address the traffic control problem with the assistance of an ML algorithm
to deal with traffic challenges. The authors employed the Q-learning RL technique for managing traffic
lights and developed an artificial environment named Simulation of Urban Mobility (SUMO) for
simulation purposes. In SUMO the cars in motion can be watched, the vehicle's delay time can be
controlled, and the delay time can be adjusted.

In [123], the aim of this paper was to set the foundation for adaptive traffic control, either by controlling
traffic lights remotely or by applying an algorithm that adjusts the timing according to the predicted flow
based on the integration of ML (RF, Linear Regression, and Stochastic Gradient Regression) and DL (MLP-
Page 18/50
NN, RNN) algorithms. The collected findings showed that the proposed ML algorithms had the worst
performance.

In [124], the authors concentrated on a critical component of ITSs known as the ability to predict lane
changes in vehicular traffic flow. The predictive accuracy to detect changes in lanes was measured using
high-fidelity data on vehicular traffic flow gathered by the US Federal Highway Administration (FHWA) for
the Peachstreet, Atlanta, GA, based on four ML models, namely SVM, NB, RF, and DT. The accuracy and
performance measurements revealed that SVM outperforms the other three ML models in terms of
precise and accurate prediction of vehicle lane shift.

In [125], a prediction approach that is based on type-2 fuzzy logic was introduced using the conceptual
framework of fuzzy logic and an urban traffic flow time series. The interval type-2 fuzzy system
prediction approach was developed, and the Back Propagation (BP) technique was utilized to update the
antecedent's coefficients and fuzzy rules' consequent. The effectiveness of the technique proposed in this
study was validated using measured data from road networks and compared to other fuzzy approaches.
The BP technique and SVM with that type-2 fuzzy logic system have a higher prediction accuracy,
according to the testing results.

In [126], the authors investigated the problem of predicting the traffic flow of a road based on historical
data. The methodology depended on the decomposition of the canonical polygonal tensor (CP) of the
traffic data. This move extracts the normal features of a traffic light on daily and weekly bases in addition
to the typical spatial allocation of traffic, while greatly minimizing the amount of data required to
represent it. Then the key elements are extended into the future, and the traffic data is regenerated from
the decomposition. The data used here is from the M62 motorway in northern England, from October 1,
2019, to October 28, 2019, at 15-minute intervals. This data is reported as the number of passing cars per
hour. Using 4 parameters, the prediction captures 90 percent of the signal's power, which exceeds the
current rolling average prediction algorithms. The authors indicated that they evaluated 4 variables in
traffic flow forecasts, but did not mention them.

In [127], the authors developed an intelligent traffic monitoring system based on ML (ML-ITMS) to
estimate traffic jams in roadside units in order to improve ITS performance. A short-term traffic flow ML-
based model was developed and SVM parameters were optimized to enhance traffic flow prediction. In
the proposed ML-ITMS, SVM and RF were specifically designed for long-range wide area networks (LoRa)
in a single query. The proposed ML-ITMS improved the accuracy estimate for traffic flow and
nonparametric processes by using mathematical models. As feedback for the proposed ML-ITMS, a data
processing method has been used. The platform was then passed through ML-ITMS services, including
public safety and security for cities, medical facility provision, traffic prediction by light and range
detection (LIDAR), and parking control. Thus, as the experimental results revealed, the proposed ML-ITMS
can improve traffic monitoring to 98.6% and can enhance traffic flow prediction systems better than other
existing methods.

Page 19/50
In [128], the authors proposed a geophysical search technique for an extreme learning machine (ELM),
called GSA-ELM. It has been suggested to unleash the performance of short-term traffic flow forecasts.
ELM avoids the cumbersome process of BP by defining the best solution analytically. The proposed
search technique generally investigates the optimal settings for ELM. The proposed search technique's
prediction performance has been measured on four standard data sets by comparing several recent
models. The four standard datasets were real-world traffic flow data from the A1, A2, A4, and A8
motorways along the Amsterdam Ring Road. The MAPEs for the GSA-ELM model on the used data sets
are 11.69%, 10.25%, 11.72%, and 12.05% respectively, while the RMSEs were 287.89, 203.04, 221.39, and
163.24, respectively.

In [129], supervised ML, as a method of Big Data analytics, to forecast various indicators of the traffic
volume were examined and conducted through two case studies. In both experiments, for training and
testing prediction models, traffic data provided by chosen automatic traffic counters on the roadways in
the Republic of Serbia, in the period from 2011 to 2018, were employed.

In [130], the authors proposed reconstructing traffic flows from the expected travel time using an ML
method. They examined the capabilities of the Gaussian Process Regressor (GPR) to handle this issue.
After obtaining the expected travel time on a specific route, a clustering method shows that travel time
profiles in each day can be associated with "different types of the day". Then, various regression factors
were trained to estimate traffic flows from the duration of travel. In this study, two situations were studied.
In the 'multi-model' variance, the regression factor was trained for each day profile. In the 'Single Model'
variation, only one Regressor was trained (the day profile was not considered). The proposed method is a
unique method to predict and reconstruct traffic flow in route networks using an ML method from
aggregated floating vehicle data (FCD). Two main problems can be identified from this work. The first
relates to using non-dispersed algorithms on the input data which can be problematic with longer
evaluation sequences, producing a more complex trained model. The other problem is a traditional issue
of every ML solution, and it has to do with the dependence on the quality of the input data.

In [131], a hybrid model incorporating ELM and ensemble-based technologies was developed to predict
the future hourly traffic on a road section in Tangiers, a city in northern Morocco. The suggested model
was built based on a high-speed ML technology that uses a kind of a Single-Layer Feed-forward Neural
Network (SLFN). The data set in this study was a set of traffic flow recorded over 5 years from 2013 to
2017 from the Moroccan Center for Road Studies and Research. This study needs to consider additional
relevant information related to traffic, such as special events, weather conditions, and traffic
characteristics on adjacent roads that may affect one particular road.

In [132], the power of various ML techniques was investigated to predict traffic conditions. Preliminary
data was collected over two weeks of monitoring in Bandung, Indonesia to be capable of determining
future traffic conditions. The collected features used in the dataset are days, hours, origins, destinations,
route view, traffic conditions, weather, and weather locations. The study investigated neural networks, NB,
DT, SVM, DNN, and DL. There are two main issues in this work. First, the size of the training data was very

Page 20/50
small. Second, the change in the training data means that the training process must be reapplied to
reflect the newer data set, which takes additional time.

In [133], the prediction accuracy of four ML models was examined using probe data gathered from the
road network of Thessaloniki, Greece. The utilized ML models were RF, Support Vector Regression (SVR),
Multilayer Perceptron (MLP), and Multiple Linear Regression (MLR). There are two key concerns in this
work. First, it has low accuracy in real-time speed prediction. Second, it needs to be tested on different
datasets.

In [7], the authors suggested a preliminary method for assessing a realistic data set of road traffic
accidents utilizing graphical representations and dimension reduction methods. The data set was
subjected to PCA analysis and linear discrimination, and the resulting performance measures provided
some comprehensive insights into the patterns of road traffic accidents. The authors developed the
preliminary framework by utilizing dimensionality reduction techniques on realistic road traffic accident
data from Gauteng Province, South Africa (SA). Furthermore, classification was carried out using the NB,
Logistic regression, and K-NN methods. The processed data was post-processed, and model performance
measures, precision, and root-mean-square error (RMSE) were used to evaluate each classifier.

In [134], the authors introduced a novel framework for stepwise regression in an idea-drift environment,
with ensemble learning as the primary solution for modernizing distribution representation. The
regression problem for predicting traffic volume was first converted into a binary classification problem.
Second, the Regression to Classification (R2C) method was used to create a more precise classification-
type loss function for ensemble learning. Finally, the regression function's incremental learning was
modeled as an incremental update to the hyper-resolution level. The proposed R2C architecture for
motion volume prediction has the disadvantage of not accounting for motion volume spatial
dependencies.

3.1.2 DL techniques for Traffic Flow Prediction

In [135], it was proposed to construct a traffic prediction system using four DL approaches namely: Deep
Autoencoder (DAN), DBN, RF, and LSTM. This technique is mostly used to estimate the traffic flow in
more populated locations. The essential parameters used in this study were zone type, weather condition,
day, road capacity, and vehicle types. There is no mention of the used dataset in this work.

In [136], the major objective was to predict trip duration from point A to point B on a route using neural
networks. Several DL and neural network algorithms were utilized such as the color clustering algorithm
(K-Means algorithm) combined with several parameters to compute and estimate travel duration. The
dataset utilized in this study was obtained using Waze Live Map APIs. The authors need to examine other
factors such as weather conditions to boost the efficiency and reliability of their job.

In [137], a short-term strategy for traffic flow forecasting based on a recurrent mixture density network,
which is a mix of RNN and mixture density network (MDN), was proposed. Traffic flow data generated by
sensors placed on road networks in Shenzhen, China, was used as the data set used in this study. It was
Page 21/50
divided into two periods: from January 1, 2019, to March 31, 2019, and from October 1, 2019, to
December 21, 2019. The modest size of the data set used is a critical issue in this study.

In [138], the authors aimed to enhance the DBN, a DL approach, performance for accurate traffic
forecasting under bad weather conditions. First, bad weather and traffic data were gathered from the IoV,
rather than from the inductance coils in the usual methods. Subsequently, the SVR technique was utilized
to improve the traditional DBN. The optimized DBN consists of two layers: the primary structure is the
traditional DBN that unsupervised learning the basic aspects of traffic data, and the topmost layer is an
SVR that implements supervised traffic forecasting. Two types of data sets were used in this study. First,
traffic data from a highway control center, and second, weather data from local monitoring stations. The
main issue in this study was that the computing time of the upgraded DBN requires optimization.

In [139], the authors proposed an urban traffic light control system that combines optimized traffic light
scheduling techniques with traffic flow forecasting techniques. The goal was to reduce the number of
vehicles that were stopped at all signal intersections on the road network. First, a framework was
proposed for an urban traffic control system, which included traffic flow predictions and signal control
optimization. Second, to alleviate traffic congestion, an interactive traffic light approach was used.
Experiments were carried out on real-world traffic data provided by the Aliyun Tianchi platform to validate
the proposed system. The comparison results showed that both the proposed system and the signal
control optimization technique work well.

In [140], the authors developed a technique for constructing a traffic congestion index by extracting free-
stream speed and flow. The author proposed the Traffic Congestion Index (TCI), which can synthesize
changes in traffic flow and speed data to assess traffic congestion, and discussed how it is generated.
Considering the correlation properties of road links in the road network, the authors introduced the
technique of grouping road links based on the sub-graph to pre-train the DL model and realize
information sharing across road links. A traffic congestion prediction model called SG-CNN was proposed
by integrating the characteristics of the traffic data and the CNN model, and the training process was
improved by the road segment aggregation method. To make the TCI more accurate, the authors have to
consider more information (such as weather, pedestrians, road conditions, etc.) that affects traffic
congestion. Furthermore, designing a more efficient algorithm while accounting for the time complexity of
the segment aggregation algorithm is an intriguing topic.

In [141], based on DL, the authors proposed a real-time data-driven queue length prediction technique.
They considered a connecting corridor on which information would be transmitted from car detectors
(placed at the intersection) to successive intersections. The length of the queue for crossing points in the
next cycle was expected to be determined by the length of the queue for the target intersection and two
upstream intersections in the current cycle. Data from the adaptive traffic control system InSync was
used to train an LSTM neural network model that extracts time-dependent patterns of a signal queue. To
reduce overfitting and to select the optimal hyperparameter combinations, the authors used a Sequential
Model-Based Optimization (SMBO) technique to determine the appropriate dropout in different stacked

Page 22/50
layers. For this investigation, they obtained adaptive traffic light data from InSync between December 18,
2017, and February 14, 2018. The Alafaya Trail (SR-434) data for East Orlando, FL, was collected from
Lake Waterford. McCulloch Road intersection includes 11 intersections. The InSync database provides
two types of data: (1) Turning Movement Counts (TMC); the number of vehicles per stage and lane per 15
minutes; (2) Historical data with details of each movement with time, duration, queue, and waiting time
for each stage. Due to the lack of data sources, it was not possible to obtain information about the
movements of the vehicles in different directions with high accuracy (30–60 seconds). If this information
is available, the performance of the model may improve further.

In [142], the authors presented AST-MTL, a multitasking learning model for predicting multi-horizon traffic
flow and velocity at the road network scale. To learn related tasks while improving generalization
performance, this approach integrates a fully connected neural network (FNN) with a multi-headed
attention mechanism. To extract the Spatio-temporal aspects of traffic states, the model incorporates
graph convolutional networks (GCNs) and GRUs. FNN begins by collecting and analyzing several related
functions in order to derive a common representation. In order to extract relevant information and
empower the model's predictive performance, the attention mechanism also considers task-specific and
shared representations. The experiments used new sets of GPS data, called OBU data, to make traffic
forecasting in highway and urban contexts. This study struggles with finding the right strategy for
explicitly maximizing task learning.

In [143], the authors proposed feature-injected RNNs (FI-RNNs), which incorporate temporal-sequential
data with contextual elements to extract the potential correlation between traffic context and state. In this
model, the stacked RNN was utilized to learn aspects of the traffic data sequence. Meanwhile, a sparse
automatic encoder has been trained to increase contextual features, which are high-level abstract
representations and coding of contextual elements. Subsequently, a fusion technique was developed that
injects contextual information into sequence features to produce fusion features. Finally, new built-in
features have been sent to the forecaster to learn traffic patterns and estimate future speed. In this study,
the accuracy and performance of the proposed model should be improved by investigating more feature
extraction and merging techniques. Also, the examination of other influencing elements is needed.

In [144], a traffic situational awareness array technology was developed, which takes advantage of
various core models. In that approach, a graph convolution was implemented on a network of traffic
detectors to extract the spatial patterns encoded in the traffic flow. After that, the retrieved features were
utilized to build a weight matrix to aggregate the predictions of the base models according to their
performance under a given condition. Traffic flow data obtained by Caltrans PeMS was used as a data
set for this study. The main observation in this study was the need to improve the network structure and
parameter options.

In [145], a traffic congestion model was proposed to predict the traffic of neighborhoods within an area
using a DL model. The model was depending on the LSTM and Graph-CNN architectures. It predicts the
degree of crowding, defined as the ratio of vehicle accumulation within a neighborhood to the trip

Page 23/50
completion rate. An abbreviated version of the San Francisco Bay Area Highway Network was used as a
data set for this study.

In [146], a strengthened Bayesian Combination Model (BCM) with DL (IBCM-DL) for traffic flow prediction
was presented to tackle the error amplification phenomenon of classical summation methods and to
improve prediction performance. The revised model was built up on the BCM framework proposed by
Wang [147]. Real-world traffic data was obtained by microwave sensors placed on highways in Beijing,
China provided the dataset for this study. Additional information, such as weather conditions, traffic
accidents, speed, and occupancy should be included to enhance the model's reliability.

In [148], the authors addressed the complexity of predicting urban traffic when an FCD is available. Four
DL methods have been compared to highlight the ability of a neural network approach (recursive and/or
convolutional) in handling the problem of traffic prediction in an urban context. In particular, the authors
investigated two RNN approaches (LSTM and GRU), as well as the spatiotemporal RCN (SRCN) model
and the High-Order Graph Convolutional LSTM Neural Network (HGC-LSTM) methods. To generate basic
FCD inputs, the proposed solutions use a traffic simulation approach. The original FCD was created with
Aimsun (2018), a microscopic traffic simulator tool for simulating each vehicle's interactions as well as
collecting data from them individually. At each pre-set period, a record (vehicle ID, speed, section, and
lane) is collected from the simulation for each associated vehicle. The assembly period was 10 seconds.
In this study, the authors evaluated the performance of prediction models using two distinct urban traffic
networks in Spain: Camp Nou, a small area of Barcelona with 4 nodes and 22 divisions, and Amara, a
district of San Sebastian with 105 nodes and 192 sections. The results of the experiments conducted
revealed that these methods can estimate traffic speeds with good performance. Specifically, recursive
algorithms (LSTM and GRU) present fewer errors than convolutional ones (SRCN and HGC-LSTM). On the
other hand, FCD can sometimes be insufficient to cover all sections of the network, and ML prediction of
a variable without any historical data is meaningless.

In [149], the authors proposed deep artificial neural network (Deep ANN) and CNN traffic speed prediction
models for upstream highway segments, including those on connected highways, under work area
conditions. The proposed models are capable of recognizing congestion on the associated links as well
as the upstream mainline segments. The suggested models predict traffic velocity under work zone
conditions based on the volume of traffic approaching the work area, speed during normal conditions,
work area capacity, distance from the work area, the vertical gradient of the road, downstream traffic
volume, and type of highway section. The proposed models utilized a dropout regulation to address the
ANN overfitting problems. The generated CNN model to predict traffic velocity under working zone
conditions should be improved in the following aspects. Discovering additional sources to update the
traffic volume to reflect the real traffic volume would enhance the accuracy of the CNN model.
Furthermore, the use of a simulation model to predict the capacity of the working area can advance the
generated CNN model. Automating databases via warehouses would facilitate the analysis of data for
new goods and developments. Additionally, provided the availability of high-resolution data, the model
can be modified to anticipate traffic congestion in the opposite direction of traffic.
Page 24/50
In [150], the authors proposed (1) an efficient and city-wide data acquisition scheme by taking a snapshot
of the Seoul Transport and Information Service (TOPIS), an open-source web-based traffic congestion
map service, and (2) by integrating CNN, LSTM, and Transpose-CNN, a hybrid neural network architecture
was created to retrieve Spatio-temporal information from the input image and predict network
congestion. In the proposed design, an LSTM network was inserted between the convolutional encoder
and the convolutional decoder. The convolutional encoder initially converts the input image sequence into
low-resolution latent state sequences. The LSTM network then learns to represent time series from
sequences, and the convolutional decoder finally converts the latent state to the original precision. To
further enhance forecast accuracy, external factors such as weather information (rain, snow, and fog)
must be addressed. Moreover, the performance of the proposed model should be enhanced. Also, more
information from many data sources must be added to get more accurate forecasts.

In [151], the authors suggested an LSTM-based traffic jam prediction technique based on correcting
missing temporal and spatial information. Before making predictions, the proposed technique performs a
pre-processing consisting of extrinsic removal using the average absolute deviation of traffic data and
correction of Spatio-temporal values ​using temporal and geographic trends and pattern data. While data
with time-series features are not effectively learned, the suggested prediction technique utilized the LSTM
model to learn time-series data to tackle this problem. The precision of forecasting traffic congestion in
low-speed areas and urban areas using the proposed technique should be enhanced. Moreover, authors
need to build a model with improved user performance.

In [152], the authors suggested a deep and embedding learning (DELA) technique that could help
explicitly learn accurate traffic information, road structure, and weather conditions. The original highway
traffic data set contained traffic flow information for approximately 3 months (from July 19, 2016, to
October 17, 2016) which was formally provided by Knowledge Discovery and Data Mining Tools
Competition (KDD CUP 2017). The proposed model has poor explanatory power for the selected DL
models. Also, it has a limited learning ability of the embed component.

In [153], an innovative and comprehensive technique for large-scale, faster, and real-time traffic
forecasting has been suggested. It has integrated four complementary advanced technologies: big data,
DL, in-memory computing, and graphics processing units (GPUs). Deep networks were trained by
employing more than 11 years of data provided by the California Department of Transportation (Caltrans)
[154]. The suggested approach has poor prediction accuracy, in addition to the use of a small size data
set.

In [155], the authors created a distinctive traffic prediction approach with the least prediction error based
on DL and introduced the LSTM model. Real-world traffic big data of performance measurement system
(PeMS) was used as the dataset of this research. The count of optimized parameters employed in this
study needs to be expanded. Also, the model training time needs to be regulated.

In [156], a pathway-based DL framework was presented. It can provide superior traffic velocity forecasts
on a citywide scale. Furthermore, the model was reasonable and interpretable in the urban transportation
Page 25/50
context. The study area was a road network consisting of 112 road sections. The dataset used was
obtained from Automated Vehicle Identification (AVI) detectors in the core area of ​Xuancheng, China.
More essential path selection criteria were investigated. Also, raising the interpretability of a DL model for
a transport application is an open topic.

In [157], using refined GPS trajectory data, the level of traffic congestion was forecasted. The Hidden
Markov model has been utilized to match GPS trajectory data to the road network. The actual speed of
road segments can even be calculated using GPS trajectory data from nearby locations. To predict
congestion levels, four DL approaches, namely CNN, RNN, LSTM, and GRU in addition to three classical
ML models (ARIMA model, SVR, and ridge regression) were used. This study had some limitations that
were highlighted. First, the GPS trajectory data collected was insufficient. Also, more GPS data must be
taken into account. In addition, the structure of the CNN network can be altered to improve model
performance.

In [158], the authors proposed a spatiotemporal model for the short-term prediction of the level of
crowding at each part of the route (CPM-ConvLSTM). The suggested model was developed on a
geographical matrix that includes both the congestion propagation pattern and the spatial correlation
between road segments. The traffic data set was obtained from Helsinki, Finland. Considering the
historical spatiotemporal matrices' time series, the authors applied the newly popular ConvLSTM DL
model by using the time series of historical spatiotemporal matrices as input and predicting the future
short-range spatiotemporal matrix. To enhance forecasting performance, the authors need to incorporate
external parameters, such as points of interest, weather, and the surrounding environment.

In [159], the authors created a DL-based methodology for directly forecasting traffic status based on a
time-space diagram using CNN. The time-space diagram is directly fed into the traffic forecasting model,
which employs a CNN. This technique has three significant benefits: (1) It allowed the time-space
diagram to be used as the input with no need for abstraction or aggregation; (2) This methodology was
created through a learning mechanism that focuses on learning the key features of the time-space
diagram required for effective forecasting. These features seriously affect the dynamic behavior of traffic
flow and vehicle interactions, which may have an impact on future traffic conditions; and (3) This
approach addressed the problem of non-parametric models' transferability by introducing location-
specific solutions that needed to be re-calibrated for another location. Compared with the existing
nonparametric models, that is, SVR, MLP, and ARIMA, the suggested CNN model provided a higher
generalization in traffic state prediction in different regions of the main diagram. The suggested CNN
model was trained using simulated data and a real-world dataset (NGSIM US-101). However, this study
did not look into the effects of lane changes on traffic flow dynamic behavior and prediction accuracy.

In [160], a new method based on fuzzy CNN (F-CNN) was proposed to predict traffic flow more accurately.
When uncertain information about traffic accidents is entered into CNN for the first time, a fuzzy
approach is used to represent traffic accident features in this method. First, to extract the Spatio-temporal
features of the traffic flow data, this study divided the whole region into 32 × 32 small blocks and created

Page 26/50
three direction sequences with inward and outward flow types. Second, by applying a fuzzy inference
mechanism, the uncertain traffic accident information was derived from the real traffic flow data. Then,
the information about the trend sequence, the information of unconfirmed traffic accidents, and the
external information can be trained by implementing the F-CNN model. Moreover, pre-training and tuning
procedures were designed efficiently to learn FCNN parameters. Finally, the Beijing taxi real route and
meteorological data sets were applied to ensure that the proposed method has superior performance
compared to the latest methods. The authors need to explore additional influential aspects in traffic flow
forecasting and use more efficient DL models.

In [161], a model for short-term traffic forecasting was proposed. This model incorporates Spatio-
temporal analysis and the GRU. Before applying an algorithm for spatiotemporal feature selection to
determine the ideal input time window and spatial data size, the proposed model applied temporal and
spatial correlation analyses to aggregated traffic flow data. Simultaneously, the desired traffic flow
information is extracted from the actual traffic flow data and converted into a two-dimensional matrix
containing Spatio-temporal traffic flow information. Finally, the GRU was employed to analyze the Spatio-
temporal features of the internal traffic flow matrix to achieve the prediction goal. There are some issues
with this work, such as other factors (for example, weather conditions) that are not included in the traffic
flow, and only the traffic flow is expected for a specific section of the road.

4. Challenges
Traffic flows must be carefully anticipated and predicted due to the risk impact of traffic congestion,
particularly in populated areas. As a result, realistic and efficient road traffic prediction techniques are
required.

The publication gap in traffic flow forecasting addressed in this survey includes a lack of computationally
effective methodologies and algorithms. Furthermore, there is a limitation of high-quality training data.
Because of using matched city traffic flow statistics, non-exhaustive data contents were used to train
network models. These characteristics were discovered to constrain the development of traffic flow
prediction using ML and DL approaches.

Because of the complicated link features between road sections and traffic congestion patterns or
congested areas, the gap is created by the underutilization of dynamically acquired Spatio-temporal
correlations in the DL. Furthermore, a lack of computing power and distributed storage constraints traffic
forecasts. A future study should look into this issue.

The current study has several limitations, including being limited to the approaches and algorithms
included in the list of articles investigated. Other strategies that were not addressed in this study could
exist. Future research should focus on popularly used DL techniques (CNN and LSTM), which are
thoroughly covered in the literature review. This is possible by using traffic data collected in various local
urban areas to provide broader data patterns for model training. As a result, traffic forecasting in small
cities will improve, as will the accuracy of the ML and DL algorithms used to predict traffic flow. The
Page 27/50
researchers' biggest challenge will be collaborating with the local urban authority to contribute the
volume of vital big data. The rules and regulations for sharing traffic data with local municipal
governments will be another impediment.

The installation of sensors to collect traffic data for training ML and DL may result in connected IoT
settings that increase cybersecurity risks. A framework should be developed to address the cybersecurity
issues of ITS in smart cities. This leaves plenty of room for future investigation.

5. Conclusion
The present study is aimed to present a comprehensive review of the most significant ML and DL
techniques used in traffic forecasting, as well as the problems associated with using ML and DL in traffic
forecasting. A total of 40 articles were chosen and thoroughly reviewed after a rigorous selection process.
According to the preceding discussion, traffic forecasting is an important task in the transportation
industry due to its significant influence on road construction, route planning, and traffic rules. This work
advances research in the field of traffic flow forecasting using ML and DL approaches. Contributes to the
literature and future studies by serving as a resource for other academics and researchers.

Declarations
Ethical Approval and Consent to participate

Not applicable.

Human and Animal Ethics

Not applicable.

Consent for publication

Not applicable.

Availability of supporting data

Not applicable.

Competing interests
Page 28/50
The authors declare that they have no known competing financial interests or personal relationships that
could have appeared to influence the work reported in this paper.

Funding

Not applicable.

Authors' contributions

Sayed A. Sayed wrote the main text of the manuscript, Yasser Abdel-Hamid and Hesham Ahmed Hefny
revised the manuscript.

Acknowledgments

Not applicable.

Authors' information

Not applicable.

References
1. Nellore K, Hancke GA (2016) Survey on Urban Traffic Management System Using Wireless Sensor
Networks. Sensors 16:157
2. Patel P, Narmawala Z, Thakkar A (2019) A survey on intelligent transportation system using internet
of things. Emerging Research in Computing, Information, Communication and Applications, 231–
240
3. An S, Lee B-H, Shin D-R (2011) A Survey of Intelligent Transportation Systems. 2011 Third
International Conference on Computational Intelligence, Communication Systems and Networks
4. Qureshi KN, Abdullah AH (2013) A Survey on Intelligent Transportation Systems. Middle-East J Sci
Res 15:629–642
5. Chen C, Li K, Teo SG, Zou X, Li K, Zeng Z (2020) Citywide traffic flow prediction based on multiple
gated Spatio-temporal convolutional neural networks. ACM Trans Knowl Discovery Data (TKDD)
14(4):1–23

Page 29/50
 6. Sun P, Boukerche A, Tao Y (2020) SSGRU: A novel hybrid stacked GRU- based traffic volume
prediction approach in a road network. Comput Commun 160:502–511
7. Makaba T, Doorsamy W, Paul BS (2020) Exploratory framework for analyzing road traffic accident
data with validation on Gauteng province data. Cogent Eng 7(1):1834659
8. World Health Organization (2018) “GLOBAL STATUS REPORT ON ROAD SAFETY 2018 SUMMARY.”
[Online]. Available: https://apps.who.int/iris/bitstream/handle/10665/277370/WHO-NMH-NVI-18.20-
eng.pdf?ua=1
9. Bengio Y (2009) Learning deep architectures for AI. Now Publishers Inc
10. Van Der Voort M, Dougherty M, Watson S (1996) Combining Kohonen maps with ARIMA time series
models to forecast traffic flow. Transp Res Part C: Emerg Technol 4(5):307–318
11. Lee S, Fambro DB (1999) Application of subset autoregressive integrated moving average model for
short-term freeway traffic volume forecasting. Transp Res Record: J Transp Res Board 1678(1):179–
188
12. Williams BM (2001) Multivariate vehicular traffic flow prediction: Evaluation of ARIMAX modeling.
Transp Res Record: J Transp Res Board 1776(1):194–200
13. Williams BM, Hoel LA (2003) Modeling and forecasting vehicular traffic flow as a seasonal ARIMA
process: Theoretical basis and empirical results. J Transp Eng 129(6):664–672
14. Chen K, Chen F, Lai B, Jin Z, Liu Y, Li K, Wei L, Wang P, Tang Y, Huang J, Hua X (2020) Dynamic
Spatio-temporal graph-based CNNs for traffic flow prediction. IEEE Access 8:185136–185145
15. Kashyap AA, Raviraj S, Devarakonda A, Nayak K, Bhat SJ (2022) Traffic flow prediction models–A
review of deep learning techniques. Cogent Eng 9(1):2010510
16. Smith BL, Demetsky MJ, &(1994) Short-term traffic flow prediction: Neural network approach.Transp.
Res. Rec,98–104
17. Simonyan K, Zisserman A(2015) Very deep convolutional networks for large-scale image recognition,
International Conference on Learning Representations, May 7–9, 2015, San Diego, USA.
https://arxiv.org/abs/1409.1556
18. Graves A, Mohamed AR, Hinton G(2013) Speech recognition with deep recurrent neural networks.
International Conference on Acoustics, Speech and Signal Processing, 26–31 May 2013, Vancouver,
Canada
19. Sainath TN, Vinyals O, Senior A, Sak H(2015) Convolutional, long short-term memory, fully connected
deep neural networks, International Conference on Acoustics, Speech and Signal Processing, 19–24
April 2015, South Brisbane, Australia
20. Good Fellow IJ, Mirza M, Courville A, Bengio Y, “Multi-prediction deep Boltzmann machines,”
Proceedings of the 26th international conference on neural information processing systems, Lake
Tahoe, USA.1
21. Sarikaya R, Hinton GE, Deoras A (2014) Application of deep belief networks for natural language
understanding. IEEE/ACM Trans Audio Speech Lang Process 22(4):778–784

Page 30/50
22. Gehring J, Miao Y, Metze F, Waibel A(2013) Extracting deep bottleneck features using stacked auto-
encoders. International Conference on Acoustics, Speech and Signal Processing, 26–31 May 2013.
IEEE, Vancouver, Canada
23. Zhang J, Wang F-Y, Wang K, Lin W-H, Xu X, Chen C (2011) Data-driven intelligent transportation
systems: a survey,. IEEE Trans Intell Transp Syst 12(4):1624–1639
24. Chowdary GJ(2021) Machine Learning and Deep Learning Methods for Building Intelligent Systems
in Medicine and Drug Discovery: A Comprehensive Survey. arXiv preprint arXiv:2107.14037
25. Singh, G., Al’Aref, S. J., Van Assen, M., Kim, T. S., van Rosendael, A., Kolli, K.K., … Min, J. K. (2018).
Machine learning in cardiac CT: basic concepts and contemporary data. Journal of Cardiovascular
Computed Tomography, 12(3), 192–201
26. Ahsan MM, Luna SA, Siddique Z (2022) March). Machine-Learning-Based Disease Diagnosis: A
Comprehensive Review. Healthcare, vol 10. MDPI, p 541. 3
27. Dey A (2016) Machine learning algorithms: a review. Int J Comput Sci Inf Technol 7(3):1174–1179
28. Dhall D, Kaur R, Juneja M(2020) Machine learning: a review of the algorithms and its applications.
Proceedings of ICRIC 2019, 47–63
29. Kotsiantis SB, Zaharakis I, Pintelas P (2007) Supervised machine learning: A review of classification
techniques. Emerg Artif Intell Appl Comput Eng 160(1):3–24
30. Veropoulos K, Campbell C, Cristianini N(1999) Controlling the Sensitivity of Support Vector
Machines. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI99)
31. Taunk K, De S, Verma S, Swetapadma A(2019) A Brief Review of Nearest Neighbor Algorithm for
Learning and Classification. 2019 International Conference on Intelligent Computing and Control
Systems (ICCS)
32. Obulesu O, Mahendra M, ThrilokReddy M(2018), July Machine learning techniques and tools: A
survey. In 2018 International Conference on Inventive Research in Computing Applications (ICIRCA)
(pp. 605–611). IEEE
33. Ray S(2019), February A quick review of machine learning algorithms. In 2019 International
conference on machine learning, big data, cloud and parallel computing (COMITCon) (pp. 35–39).
IEEE
34. Kumar R, Verma R (2012) Classification algorithms for data mining: A survey. Int J Innovations Eng
Technol (IJIET) 1(2):7–14
35. Nikam SS (2015) A comparative study of classification techniques in data mining algorithms. Orient
J Comput Sci Technol 8(1):13–19
36. Stein G, Chen B, Wu AS, Hua KA(2005) “Decision tree classifier for network intrusion detection with
GA-based feature selection,” in Proceedings of the 43rd annual Southeast regional conference-
Volume 2, pp. 136–141
37. Damanik IS, Windarto AP, Wanto A, Andani SR, Saputra W(2019) “Decision Tree Optimization in C4. 5
Algorithm Using Genetic Algorithm,” in Journal of Physics: Conference Series, vol. 1255, no. 1,

Page 31/50
 p. 012012
38. Gavankar SS, Sawarkar SD(2017) "Eager decision tree", in 2017 2nd International Conference for
Convergence in Technology (I2CT), Mumbai, Apr. pp. 837–840
39. Mahesh B (2020) Machine learning algorithms-a review. Int J Sci Res (IJSR) [Internet] 9:381–386
40. Janikow CZ(1998) "Fuzzy decision trees: issues and methods," IEEE Transactions on Systems, Man,
and Cybernetics, Part B (Cybernetics), vol. 28, no. 1, pp. 1–14,
41. Charbuty B, Abdulazeez A (2021) Classification based on decision tree algorithm for machine
learning. J Appl Sci Technol Trends 2(01):20–28
42. Zhao Y, Zhang Y (2008) Comparison of decision tree methods for finding active objects. Adv Space
Res 41(12):1955–1959
43. Mittal K, Khanduja D, Tewari PC (2017) An insight into ‘Decision Tree Analysis’. World Wide Journal
of Multidisciplinary Research and Development 3(12):111–115
44. Priyanka, Kumar D (2020) Decision tree classifier: a detailed survey. Int J Inform Decis Sci
12(3):246–269
45. Breiman L (2001) “Random Forests ” Machine Learning 54(1):5–32
46. Pal M (2005) Random Forest Classifier for Remote Sensing Classification. Int J Remote Sens
26(1):217–222
47. Cutler DR, Edwards TC Jr, Beard KH, Cutler A, Hess KT, Gibson J, Lawler JJ (2007) Random Forests
for Classification in Ecology. Ecology 88(11):2783–2792
48. Belgiu M, Drăguţ L (2016) Random Forest in Remote Sensing: A Review of Applications and Future
Directions. ISPRS J Photogrammetry Remote Sens 114:24–31
49. He Y, Lee E, Warner TA (2017) A Time Series of Annual Land Use and Land Cover Maps of China
from 1982 to 2013 Generated Using AVHRR GIMMS NDVI3g Data. Remote Sens Environ 199:201–
217
50. Maxwell AE, Warner TA, Fang F (2018) Implementation of machine-learning classification in remote
sensing: An applied review. Int J Remote Sens 39(9):2784–2817
51. Breiman(2001) Random forests. Machine Learning 45, 1 (Oct. 2001), 5–32
52. Yali Amit and Donald Geman (1997) Shape quantization and recognition with randomized trees.
Neural Computation 9, 7 (1997), 1545–1588
53. Tin KH(1995) Random decision forests. In 3rd International Conference on Document Analysis and
Recognition - Volume 1 (ICDAR’95). IEEE Computer Society, 278–282
54. Tin Kam Ho (1998) The random subspace method for constructing decision forests. 20, 8 (Aug.
1998), 832–844
55. Resende PAA, Drummond AC (2018) A survey of random forest-based methods for intrusion
detection systems. ACM Comput Surv (CSUR) 51(3):1–36
56. Oladipo ID, AbdulRaheem M, Awotunde JB, Bhoi AK, Adeniyi EA, Abiodun MK (2022) Machine
Learning and Deep Learning Algorithms for Smart Cities: A Start-of-the-Art Review. IoT and IoE Driven
Page 32/50
 Smart Cities, pp 143–162
57. Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., … Lautenbach,S. (2013).
Collinearity: a review of methods to deal with it and a simulation study evaluating their performance.
Ecography, 36(1), 27–46
58. Harrington P (2012) Machine Learning in action. Manning Publications Co., Shelter Island, New York
59. McLachlan GJ (2005) Discriminant analysis and statistical pattern recognition. John Wiley & Sons
60. Jolliffe IT (1986) Principal Component Analysis. SpringerVerlag, New York
61. Gow J, Baumgarten R, Cairns P, Colton S, Miller P (2012) Unsupervised modeling of player style with
LDA. IEEE Trans Comput Intell AI Games 4(3):152–166
62. Coronato A, Naeem M, De Pietro G, Paragliola G (2020) Reinforcement learning for intelligent
healthcare applications: A survey. Artif Intell Med 109:101964
63. Watkin CJCH, Dayan P (1992) Technical note Q-learning. Mach Learn 8(3):279–292
64. Watkins CJCH(1989) Learning from delayed rewards. Ph.D. Thesis, University of Cambridge, England
65. Achille A, Soatto S(2018) "Information dropout: Learning optimal representations through noisy
computation,"IEEE Transactions on Pattern Analysis and Machine Intelligence,
66. Williams G, Wagener N, Goldfain B, Drews P, Rehg JM, Boots B, Theodorou EA(2017) "Information-
theoretic mpc for model-based reinforcement learning," In Robotics and Automation (ICRA), IEEE
International Conference on, pp. 1714–1721,
67. Wilkes JT, Gallistel CR(2017) "Information theory, memory, prediction, and timing in associative
learning," Computational Models of Brain and Behavior, pp. 481–492,
68. Jang B, Kim M, Harerimana G, Kim JW (2019) Q-learning algorithms: A comprehensive classification
and applications. IEEE Access 7:133653–133667
69. An Y, Wang Y, Meng H(2017) "Multi-task deep learning for user intention understanding in speech
interaction systems,"
70. Shi X, Gao Z, Lausen L, Wang H, Yeung DY, Wong WK, Woo WC(2017) "Deep learning for precipitation
nowcasting: A benchmark and a new model," Advances in Neural Information Processing Systems,"
pp.5622–5632,
71. Chia-Feng Juang and Chun-Ming, Lu(2009) Ant colony optimization incorporated with fuzzy Q-
learning for reinforcement fuzzy control. IEEE Transactions on Systems, Man, and Cybernetics-Part
A: Systems and Humans 39, 3 (2009), 597–608
72. Świechowski M, Godlewski K, Sawicki B, Mańdziuk J(2021) Monte Carlo tree search: A review of
recent modifications and applications.arXiv preprint arXiv:2103.04931
73. Lizotte DJ, Laber EB (2016) Multi-Objective Markov Decision Processes for Data-Driven Decision
Support. J Mach Learn Res 17:211:1–211
74. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, Van Den Driessche G, Schrittwieser J, Antonoglou I,
Panneershelvam V, Lanctot M et al (2016) Mastering the Game of Go with Deep Neural Networks and
Tree Search. Nature 529(7587):484–489
Page 33/50
75. Browne, C. B., Powley, E., Whitehouse, D., Lucas, S. M., Cowling, P. I., Rohlfshagen,P., … Colton, S.
(2012). A survey of monte carlo tree search methods. IEEE Transactions on Computational
Intelligence and AI in games, 4(1), 1–43
76. Baier H, Drake PD(2010) “The power of forgetting: Improving the last good-reply policy in Monte
Carlo Go,” IEEE Trans. Comput. Intell. AI Games, vol. 2, no. 4, pp. 303–309, Dec.
77. Alpaydin E (2020) Introduction to machine learning. MIT press
78. Mikolov T et al(2013) Efficient estimation of word representations in vector space.
79. Nguyen G, Dlugolinsky S, Bobák M, Tran V, García ÁL, Heredia I et al (2019) Machine learning and
deep learning frameworks and libraries for large-scale data mining: a survey. Artif Intell Rev
52(1):77–124
80. Aggour KS, Gupta VK, Ruscitto D, Ajdelsztajn L, Bian X, Brosnan KH et al (2019) Artificial
intelligence/machine learning in manufacturing and inspection: a GE perspective. MRS Bull
44(7):545–558
81. Khan FN, Fan Q, Lu C, Lau APT (2020) Machine learning methods for optical communication
systems and networks. Optical fiber telecommunications VII. Academic Press, New York, pp 921–978
82. Pouyanfar S, Sadiq S, Yan Y, Tian H, Tao Y, Reyes MP et al (2018) A survey on deep learning:
algorithms, techniques, and applications. ACM Comput Surv (CSUR) 51(5):1–36
83. Dargan S, Kumar M, Ayyagari MR, Kumar G (2019) A survey of deep learning and its applications: a
new paradigm to machine learning. Arch Comput Methods Eng 27(4):1–22
84. Lauzon FQ(2012) An introduction to deep learning, In 2012 11th International Conference on
Information Science, Signal Processing and their Applications (ISSPA), IEEE. pp. 1438–1439
85. Ling, Z. H., Kang, S. Y., Zen, H., Senior, A., Schuster, M., Qian, X. J., … Deng,L. (2015). Deep learning for
acoustic modeling in parametric speech generation: A systematic review of existing techniques and
future trends. IEEE Signal Processing Magazine, 32(3), 35–52
86. Schmidhuber J (2015) Deep learning in neural networks: An overview. Neural Netw 61:85–117
87. Yu D, Deng L (2010) Deep learning and its applications to signal and information processing
[exploratory dsp]. IEEE Signal Process Mag 28(1):145–154
88. Yap, M. H., Pons, G., Martí, J., Ganau, S., Sentis, M., Zwiggelaar, R., … Marti, R.(2017). Automated
breast ultrasound lesions detection using convolutional neural networks.IEEE Journal of biomedical
and health informatics, 22(4), 1218–1226
89. Fukushima K, Neocognitron (1988) A hierarchical neural network capable of visual pattern
recognition. Neural Netw 1:119–130
90. LeCun Y, Bottou L, Bengio Y, Haffner P(1998) Gradient-based learning applied to document
recognition. Proc. IEEE 86, 2278–2324
91. Goodfellow I, Bengio Y, Courville A(2016)Deep Learning; MIT Press: Cambridge, MA, USA,
92. Alom, M. Z., Taha, T. M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M. S., …Asari, V. K. (2019). A
state-of-the-art survey on deep learning theory and architectures.Electronics, 8(3), 292

Page 34/50
 93. Apaydin H, Feizi H, Sattari MT, Colak MS, Shamshirband S, Chau KW (2020) Comparative analysis of
recurrent neural network architectures for reservoir inflow forecasting. Water 12(5):1500
94. Graves A, Mohamed AR, Hinton G(2013), May Speech recognition with deep recurrent neural
networks. In 2013 IEEE international conference on acoustics, speech and signal processing
(pp. 6645–6649)
95. BATUR DİNLER Ö, Aydin N (2020) An optimal feature parameter set based on gated recurrent unit
recurrent neural networks for speech segment detection. Appl Sci 10(4):1273
96. Jagannatha AN, Yu H(2016), November Structured prediction models for RNN-based sequence
labeling in clinical text. In Proceedings of the conference on empirical methods in natural language
processing. conference on empirical methods in natural language processing (Vol. 2016, p. 856). NIH
Public Access
97. Alzubaidi, L., Zhang, J., Humaidi, A. J., Al-Dujaili, A., Duan, Y., Al-Shamma, O.,… Farhan, L. (2021).
Review of deep learning: Concepts, CNN architectures, challenges,applications, future directions.
Journal of Big Data, 8(1), 1–74
98. Pascanu R, Gulcehre C, Cho K, Bengio Y(2013) How to construct deep recurrent neural networks.arXiv
preprint arXiv:1312.6026
99. Glorot X, Bengio Y(2010) Understanding the difficulty of training deep feedforward neural networks.
In: Proceedings of the thirteenth international conference on artificial intelligence and statistics;
p. 249–56
100. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780
101. Smagulova K, James AP (2019) A survey on LSTM memristive neural network architectures and
applications. Eur Phys J Special Top 228(10):2313–2324
102. Setyanto, A., Laksito, A., Alarfaj, F., Alreshoodi, M., Oyong, I., Hayaty, M., … Kurniasari,L. (2022). Arabic
Language Opinion Mining Based on Long Short-Term Memory (LSTM).Applied Sciences, 12(9), 4140
103. Gers FA, Schmidhuber J, Cummins F (2000) Learning to forget: Continual prediction with LSTM.
Neural Comput 12(10):2451–2471
104. Lindemann B, Müller T, Vietz H, Jazdi N, Weyrich M (2021) A survey on long short-term memory
networks for time series prediction. Procedia CIRP 99:650–655
105. Cui Z, Ke R, Pu Z, Wang Y(2018) Deep bidirectional and unidirectional LSTM recurrent neural network
for network-wide traffic speed prediction.arXiv preprint arXiv:1801.02143
106. Villegas R, Yang J, Zou Y, Sohn S, Lin X, Lee H(2017), July Learning to generate long-term future via
hierarchical predictin. In international conference on machine learning (pp. 3560–3569). PMLR
107. Chu KF, Lam AY, Li VO (2019) Deep multi-scale convolutional LSTM network for travel demand and
origin-destination predictions. IEEE Trans Intell Transp Syst 21(8):3219–3232
108. Gensler A, Henze J, Sick B, Raabe N(2016), October Deep Learning for solar power forecasting—An
approach using AutoEncoder and LSTM Neural Networks. In 2016 IEEE international conference on
systems, man, and cybernetics (SMC) (pp. 002858–002865). IEEE

Page 35/50
109. Hsu D(2017) Multi-period time series modeling with sparsity via Bayesian variational inference. arXiv
preprint arXiv:1707.00666.
110. Kalchbrenner N, Danihelka I, Graves A(2015) Grid long short-term memory. arXiv preprint
arXiv:1507.01526.
111. Veličković, P., Karazija, L., Lane, N. D., Bhattacharya, S., Liberis, E., Liò, P.,… Vegreville, M. (2018, May).
Cross-modal recurrent models for weight objective prediction from multimodal time-series data. In
Proceedings of the 12th EAI International Conference on Pervasive Computing Technologies for
Healthcare (pp. 178–186)
112. Wang J, Hu X (2021) Convolutional neural networks with gated recurrent connections. IEEE
Transactions on Pattern Analysis and Machine Intelligence
113. Liang M, Hu X(2015) Recurrent convolutional neural network for object recognition. In Proceedings of
the IEEE conference on computer vision and pattern recognition (pp. 3367–3375)
114. Liang M, Hu X, Zhang B(2015) Convolutional neural networks with intra-layer recurrent connections
for scene labeling.Advances in neural information processing systems,28
115. Fernandez B, Parlos AG, Tsai W(1990) Nonlinear dynamic system identification using artificial neural
networks (anns). In International Joint Conference on Neural Networks (IJCNN), pages 133–141,
116. Puskorius GV, Feldkamp LA (1994) Neurocontrol of nonlinear dynamical systems with Kalman filter
trained recurrent networks. IEEE Trans Neural Networks 5(2):279–297
117. Rumelhart DE, Hinton GE, Williams RJ (1986) Parallel distributed processing: Explorations in the
microstructure of cognition. chapter Learning Internal Representations by Error Propagation, vol 1.
MIT Press, pp 318–362
118. Lippi M, Bertini M, Frasconi P (2013) Short-term traffic flow forecasting: An experimental comparison
of time-series analysis and supervised learning. IEEE Trans Intell Transp Syst 14(2):871–882
119. Aqib M, Mehmood R, Alzahrani A, Katib I, Albeshri A, Altowaijri SM (2019) Smarter Traffic Prediction
Using Big Data, In-Memory Computing, Deep Learning and GPUs. Sensors 19:2206
120. Janković S, Uzelac A, Zdravković S, Mladenović D, Mladenović S, Andrijanić I(2021) TRAFFIC
VOLUMES PREDICTION USING BIG DATA ANALYTICS METHODS.International Journal for Traffic &
Transport Engineering, 11(2)
121. Deekshetha HR, Madhav S, Tyagi AK (2022) Traffic Prediction Using Machine Learning. Evolutionary
Computing and Mobile Sustainable Networks. Springer, Singapore, pp 969–983
122. Suneel Kuamr (2022) Traffic Flow Prediction Using Machine Learning Algorithms. Int Res J Eng
Technol (IRJET) 9(4):2995–3004
123. Navarro-Espinoza A, López-Bonilla OR, García-Guerrero EE, Tlelo-Cuautle E, López-Mancilla D,
Hernández-Mejía C, Inzunza-González E (2022) Traffic Flow Prediction for Smart Traffic Lights Using
Machine Learning Algorithms. Technologies 10(1):5
124. Upadhyaya S, Mehrotra D(2022) The Facets of Machine Learning in Lane Change Prediction of
Vehicular Traffic Flow. In Proceedings of International Conference on Intelligent Cyber-Physical

Page 36/50
 Systems (pp. 353–365). Springer, Singapore
125. Qu Z, Li J(2022) Short-term Traffic Flow Forecast on Basis of PCA-Interval Type-2 Fuzzy System. In
Journal of Physics: Conference Series (Vol. 2171, No. 1, p. 012051). IOP Publishing
126. Steffen T, Lichtenberg G (2022) February). A Machine Learning Approach to Traffic Flow Prediction
using CP Data Tensor Decompositions. IFAC World Congress 2020. Loughborough Research
Repository
127. Wang J, Pradhan MR, Gunasekaran N (2022) Machine learning-based human-robot interaction in
ITS. Inf Process Manag 59(1):102750
128. Cui Z, Huang B, Dou H, Tan G, Zheng S, Zhou T (2022) GSA-ELM: A hybrid learning model for short‐
term traffic flow forecasting. IET Intel Transport Syst 16(1):41–52
129. Janković S, Uzelac A, Zdravković S, Mladenović D, Mladenović S, Andrijanić I(2021) TRAFFIC
VOLUMES PREDICTION USING BIG DATA ANALYTICS METHODS.International Journal for Traffic &
Transport Engineering, 11(2)
130. Li J, Boonaert J, Doniec A, Lozenguez G (2021) Multi-models machine learning methods for traffic
flow estimation from Floating Car Data. Transp Res Part C: Emerg Technol 132:103389
131. Jiber M, Mbarek A, Yahyaouy A, Sabri MA, Boumhidi J (2020) Road traffic prediction model using
extreme learning machine: the case study of Tangier, Morocco. Information 11(12):542
132. Husni E, Nasution SM, Yusuf R (2020) Predicting Traffic Conditions Using Knowledge-Growing Bayes
Classifier. IEEE Access 8:191510–191518
133. Bratsas C, Koupidis K, Salanova JM, Giannakopoulos K, Kaloudis A, Aifadopoulou G (2020) A
comparison of machine learning methods for the prediction of traffic speed in urban places.
Sustainability 12(1):142
134. Xiao J, Xiao Z, Wang D, Bai J, Havyarimana V, Zeng F (2019) Short-term traffic volume prediction by
ensemble learning in concept drifting environments. Knowl Based Syst 164:213–225
135. Nazirkar R, Ramchandra C, Rajabhushanam(2021) Machine learning algorithms performance
evaluation in traffic flow prediction, Materials Today: Proceedings, ISSN 2214–7853
136. Pangesta J, Dharmadinata OJ, Bagaskoro ASC, Hendrikson N, Budiharto W (2021) Travel duration
prediction based on traffic speed and driving pattern using deep learning. ICIC express letters. Part b.
applications: an international journal of research and surveys 12(1):83–90
137. Chen M, Chen R, Cai F, Li W, Guo N, Li G(2021) Short-Term Traffic Flow Prediction with Recurrent
Mixture Density Network. Mathematical Problems in Engineering, 2021
138. Bao X, Jiang D, Yang X, Wang H (2021) An improved deep belief network for traffic prediction
considering weather factors. Alexandria Eng J 60(1):413–420
139. Jiang CY, Hu XM, Chen WN(2021), May An Urban Traffic Signal Control System Based on Traffic Flow
Prediction. In 2021 13th International Conference on Advanced Computational Intelligence (ICACI)
(pp. 259–265). IEEE

Page 37/50
140. Tu Y, Lin S, Qiao J, Liu B (2021) Deep traffic congestion prediction model based on road segment
grouping. Appl Intell 51(11):8519–8541
141. Rahman R, Hasan S (2021) Real-time signal queue length prediction using long short-term memory
neural network. Neural Comput Appl 33(8):3311–3324
142. Buroni G, Lebichot B, Bontempi G (2021) AST-MTL: An Attention-Based Multi-Task Learning Strategy
for Traffic Forecasting. IEEE Access 9:77359–77370
143. Qu L, Lyu J, Li W, Ma D, Fan H (2021) Features injected recurrent neural networks for short-term traffic
speed prediction. Neurocomputing 451:290–304
144. Chen Y, Lv Y, Ye P, Zhu F (2020) Traffic-Condition-Awareness Ensemble Learning for Traffic Flow
Prediction. IFAC-PapersOnLine 53(5):582–587
145. Mohanty S, Pozdnukhov A, Cassidy M (2020) Region-wide congestion prediction and control using
deep learning. Transp Res Part C: Emerg Technol 116:102624
146. Gu Y, Lu W, Xu X, Qin L, Shao Z, Zhang H (2020) An improved Bayesian combination model for short-
term traffic prediction with deep learning. IEEE Trans Intell Transp Syst 21(3):1332–1342
147. Wang J, Deng W, Guo Y(2014) “New Bayesian combination method for short-term traffic flow
forecasting,” Transp. Res. C, Emerg. Technol., vol. 43, pp. 79–94, Jun.
148. Vázquez JJ, Arjona J, Linares M, Casanovas-Garcia J (2020) A comparison of deep learning
methods for urban traffic forecasting using floating car data. Transp Res Procedia 47:195–202
149. Shabarek A(2020) A Deep Machine Learning Approach for Predicting Freeway Work Zone Delay
Using Big Data (Doctoral dissertation, New Jersey Institute of Technology)
150. Ranjan N, Bhandari S, Zhao HP, Kim H, Khan P (2020) City-wide traffic congestion prediction based
on CNN, LSTM, and transpose CNN. IEEE Access 8:81606–81620
151. Shin DH, Chung K, Park RC (2020) Prediction of traffic congestion based on LSTM through correction
of missing temporal and spatial data. IEEE Access 8:150784–150796
152. Zheng Z, Yang Y, Liu J, Dai HN, Zhang Y (2019) Deep and embedded learning approach for traffic
flow prediction in urban informatics. IEEE Trans Intell Transp Syst 20(10):3927–3939
153. Aqib M, Mehmood R, Alzahrani A, Katib I, Albeshri A, Altowaijri SM (2019) Smarter Traffic Prediction
Using Big Data, In-Memory Computing, Deep Learning and GPUs. Sensors 19:2206
154. California Department of Transportation (Caltrans) (2019) Caltrans Performance Measurement
System (PeMS) Available online: http://pems.dot.ca.gov/ (accessed on 13
155. Kong F, Li J, Jiang B, Zhang T, Song H(2019) Big data-driven machine learning‐enabled traffic flow
prediction.Transactions on Emerging Telecommunications Technologies, 30(9), e3482
156. Wang J, Chen R, He Z (2019) Traffic speed prediction for urban transportation network: A path based
deep learning approach. Transp Res Part C: Emerg Technol 100:372–385
157. Sun S, Chen J, Sun J (2019) Traffic congestion prediction based on GPS trajectory data. Int J Distrib
Sens Netw 15(5):1550147719847440

Page 38/50
158. Di X, Xiao Y, Zhu C, Deng Y, Zhao Q, Rao W(2019), June Traffic congestion prediction by
spatiotemporal propagation patterns. In 2019 20th IEEE International Conference on Mobile Data
Management (MDM) (pp. 298–303). IEEE
159. Khajeh Hosseini M, Talebpour A (2019) Traffic prediction using time-space diagram: A convolutional
neural network approach. Transp Res Rec 2673(7):425–435
160. An J, Fu L, Hu M, Chen W, Zhan J (2019) A novel fuzzy-based convolutional neural network method
to traffic flow prediction with uncertain traffic accident information. IEEE Access 7:20708–20722
161. Dai G, Ma C, Xu X (2019) Short-term traffic flow prediction method for urban road sections based on
space-time analysis and GRU. IEEE Access 7:143025–143035

Figures

Figure 1

Traffic stakeholders

Page 39/50
Figure 2

Benefits of traffic flow forecasting

Figure 3

AI, ML, and DL

Page 40/50
Figure 4

Various types of ML algorithms [26]

Page 41/50
Figure 5

A hyperplane separating two classes [27]

Figure 6

Maximum Margin [29]

Page 42/50
Figure 7

DT Structure [40]

Figure 8

Unsupervised learning workflow [28].

Page 43/50
Figure 9

Data visualization before and after applying PCA [58].

Figure 10

The basic MCTS process [76]

Page 44/50
Figure 11

ML vs DL.

Figure 12

ML and DL algorithms development timeline [26]

Page 45/50
Figure 13

A CNN architecture.

Figure 14

Feature extractors and classifier parts of CNN [92].

Page 46/50
Figure 15

A typical unfolded RNN diagram [97].

Figure 16

Page 47/50
(a) original LSTM cell architecture; (b) LSTM cell includes forget gate [101].

Figure 17

(left) Bidirectional LSTM and (right) filter mechanism for processing incomplete data sets [105].

Figure 18

Convolution operations within LSTM cells [107]

Page 48/50
Figure 19

Grid LSTM (right) vs Stacked LSTM (left) [110]

Figure 20

Cross-modal LSTM [111]

Page 49/50
Figure 21

The CNN, RMLP, and RCNN architectures [113].

Figure 22

RCNN with one convolutional layer, four RCLs, three max-pooling layers, and one softmax layer [113]

Page 50/50