Multimedia Tools and Applications

https://doi.org/10.1007/s11042-024-20364-y

A hybrid model for detecting e‑commerce product returns

using CNN‑LSTM

Vidya Rajasekaran1 · Latha Tamilselvan2

Received: 27 September 2023 / Revised: 27 June 2024 / Accepted: 1 October 2024

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024

Abstract
This paper aims to detect the return of products using deep learning techniques. We work
on generating a score for every product sold through the e-commerce firm using the con-
sumer’s feedback and their ratings. An automated system for identifying the worth of every
product will help the e-commerce firm to detect the products with negative reviews and
low ratings which will drop down the fame of the e-commerce and in turn may increase
the number of product returns to the origin. This defames and product returns may reflect
in reducing the revenue of the firm. In this work, we developed a novel hybrid classifi-
cation model combining Recurrent Neural Network (RNN) along with Long Short-Term
Memory (LSTM). Hybrid deep learning model that combines LSTM networks and RRN
is built and tested on the dataset. The results of the hybrid model are compared against
four individual models Naive Bayes, SVM, LSTM and RNN. The model is developed and
evaluated in python using keras and tensor flow as backend. The developed hybrid model
outperforms with an increased accuracy of 7% to 8% compared to the individual LSTM
and CNN models.

Keywords Deep learning · E-commerce · Novel hybrid · LSTM and CNN models · RNN

1 Introduction

Internet users who shop through online e-commerce sites are rising every year. The ongoing
digitalization increases the number of digital buyers. With this exponential increase of online
purchases, they also contribute to a notable amount of return issues. The RTO (Return to Ori-
gin) is a major issue which affects the overall revenue of the e-commerce firm. They in turn
increase the financial and operational losses. Every individual customer needs to be considered

* Latha Tamilselvan
latha.tamil@crescent.education
Vidya Rajasekaran
vidyarajesh23@gmail.com
1
Department of CSE (AIML), St. Peter’s Engineering College, Hyderabad, India
2
Department of Information Technology, B.S. Abdur Rahman Crescent Institute of Science
and Technology, Chennai, India

Vol.:(0123456789)
 Multimedia Tools and Applications

as a valuable asset for the e-commerce firm. So, the firm needs to handle RTO issues with
care, which never harm the consumers, nor the vendors and the revenue of the firm. So, devel-
oping a detection and prediction model to solve the return issues will enhance the revenue of
the firm. Detecting the real cause for returns is a major challenge and providing a solution for
the cause will greatly reduce the number of returns and prevent the unnecessary losses. Only
satisfied customers will make no returns and will purchase again with the e-commerce firm,
which will reduce the customer churn and altogether these will improve the revenue of the
firm. The reason for return needs to be clearly investigated. The reason can fall into any cate-
gory like quality issues with the product, product difference from displayed image and original
received one, delay in delivery time, damaged product and even due to the fraudulent behavior
of the customers. So, the detection and prediction model should be developed with most care
to properly investigate the real cause. All this can be studied from the consumer’s feedbacks
and their ratings along with the product sales and consumer’s transaction history. So, we focus
our research and work on consumer, product and transaction data and build models to make
detection and predictions. Evaluating and analysis of the above-mentioned data can make val-
uable contributions towards revenue generation. A huge volume of the consumer transaction,
product sales and consumer feedback are generated by the e-commerce firm. The development
of machine learning algorithms helps in processing those voluminous data and extracts the
valuable insights from it. Knowledge of the data can be extracted with the developed model
and analytical techniques can be applied to it.
LSTM is widely used in several prediction problems and it surmounts vanishing gra-
dient problem occurred in classic RNN (Recurrent Neural Network). LSTM suffers from
exploding gradient problem [1, 2].
CNN is suited for text classification with good accuracy because of the non-linearity of
the network and also it easily integrates the pre-trained word embeddings. But the Maxpool
layer slows down the operation of CNN, and for CNNs with several layers, the training pro-
cess takes more time [3–5].
A hybrid CNN-LSTM model with improved accuracy is developed based on the
reviews. There are several CNN-LSTM combined models applied for different applications
[6–8]. In general, the CNN-LSTM model is an arrangement of CNN layers that extort the
features from the input data and LSTM layers to produce the sequential predictions [9, 10].
The very basic model in Fig. 1 represents the classical LSTM.
The CNN network as shown in Fig. 0.2.is developed using two Convolutional layers and
used for classifying the products to be returned or not. It takes the word embedded as input,
assigns weights and biases, and performs classification. CNN performs faster when com-
pared to RNN, because CNN were basically designed for images and RNN for texts, and
also CNN does not build the relationship between hidden vectors for each step.

2 Related research

2.1 Opinion mining

Opinion mining or sentiment analysis is the process of extracting the emotions from the body
of the text, using NLP (Natural Language Processing) techniques [11, 12]. Sentiment analysis
can be applied for classifying the product review and develop recommendations so that other
consumers can decide their purchase based on the previous reviews [13]. In [14] to understand
the service quality sentiment analysis is been used. The better the services the better the positive
Multimedia Tools and Applications

Fig. 1  Classical LSTM

sentiments and is vice versa for negative sentiments. The analysis made by [15] states that LSTM
and CNN algorithms are the two most used ones in sentiment analysis. The paper [16] illustrates
how deep learning can be applied to analyze the data on a large scale for the real time applica-
tions such as e-commerce, health care, agriculture and many others. The work [17] reviews the
applications of NLP in business. The researches work in [18] analysis the consumer’s behav-
ior and its importance to develop the e-commerce marketing strategy. The correlation between
the review and the product recommendation were found using NLP techniques. So, the main
idea was to study the consumer psychology which can be further improved and develop new
strategies.

2.2 Classic LSTM model

LSTM is a sophisticated form of recurrent neural network (RNN) that not only works on
solitary data points but also on complete sequences of the data. The LSTM in general per-
forms better flow of Information through the unit. LSTM overcomes the vanishing gradient
problem of classic RRN but suffers from an exploding gradient problem when compared to
vanilla RNN. LSTM provides more flexibility in controlling the outputs and yields better
results. The structure of LSTM consists of input, output and forget gate, which helps solve
the long-term data(dependency problems of traditional RNN problems.
The forget gate ft representation is shown in Eq. (1),
)

(1)
( ) ( )
ft = 𝜎g wf xt + uf ht−1 + bf

The input function it is defined as in Eq. (2),

(2)
( ) ( )
it = 𝜎g wi xt + ui ht−1 + bi

The output gate produces the output and is represented as shown in Eq. (3),
 Multimedia Tools and Applications

( ) ( )
ot = 𝜎 g w0 xt + u0 ht−1 + b0 (3)

Here 𝜎g is the sigmoid function.wg and uz are the weights of the inputs and recurrent con-
nections that are used to solve the vanishing gradient problem. Z can be either the input (i),
the output (o) or forget (f) gate. It can also be cell memory(c) which is based on the activation
computed. The control gate ct is used to update the memory cell and is represented as in Eq. (4),

(4)
( ) (( ) ( ))
ct = ft ×ct−1 + it ×̃at

where,

(5)
( ) ( )
ã t = 𝜎c wc xt + uc ht−1 + bc

And

(6)
( ) ( ( ))
ht = ot × 𝜎h ct

Here 𝜎h is the tangent function, and c0 = 0;h0 = 0.The sequential layer requires a large
amount of memory bandwidth for computation which affects the performance of the out-
put. The results are shown in the Table 1.

2.3 CNN model (convolutional neural network)

CNNs are usually used for classifying images and videos. There are research works
[19–21], where CNN were applied for NLP and resulted with enhanced performance. So
we tried applying CNN also as shown in Fig. 2.

3 Empirical setting

3.1 Dataset description

The e-commerce consumer’s feedback dataset is used from the source (https://​www.​kag-
gle.​com/​datas​ets/​dataf​i niti/​consu​mer-​revie​ws-​of-​amazon-​produ​cts) with 34,000 reviews
from the customers for different products sold on the e-commerce site. For our experi-
mental analysis we carried out data processing with the original dataset. We performed
data quality assessment to check the mismatched data types, any mixed data values occur,
data outliers available or any missing data is found. Then move on to data cleaning to add
the missing values and removing / repairing the irrelevant data in the dataset. Next Data
transformation were carried out to turn the data into proper format make analysis. Feature
selection and normalization techniques were applied. Feature selection is the process of
selecting the features to train our ML models using RFE (Recursive Feature Elimination).
Normalization was used to scale the data within a specified range. Data reduction is carried
out to reduce the amount of data used to facilitate the analysis and downstream the process.
The final retrieved dataset applied for the research work is shown in the Fig. 3.
Multimedia Tools and Applications

Fig. 2  Architecture of CNN-

based model for product returns
prediction

3.2 Variable description

The proposed work is carried out using seven variables as input parameters and one
variable as output parameter. The learning function X denotes input variables and Y
denotes output variable as shown in (7).
Y = f (X) (7)
 Multimedia Tools and Applications

Fig. 3  Sample Dataset Used for Learning

The data is split into input and output variables using the slice operator.
The variables used for prediction of product return are explored and defined as (X)
and the response variable as (Y). In our research we use several predictor variables as
denoted in Table 1 and their respective descriptions are also stated along with.

3.3 Evaluation metrics

The resulting performance of the used models were monitored and measured using the
following metrics as represented in Eqs. (8) through (11):

• Accuracy (ACC)
• Sensitivity (SN)
• Specificity (SP)
• Precision (P).

The Accuracy is calculated as the ratio between the numbers of correct predictions
made to the total number of predictions done.

Table 1  Illustration of variables

Category Variable Names Description

Input Variable (X) Product_id Unique product identification number

Reviews Ratings
Numeric rating in the range between 1 to 5
Reviews Text Feedback in textual format from the consumers about the
product
Reviews Title Title reflecting the feedback
Reviews username The consumers identification name writing reviews which is not
unique
Return habit The attitude of making returns by the consumer
Polarity score The emotion on the feedback either positive or negative about
the product
Output Variable (Y) Prediction Score The score representing the return of product
Multimedia Tools and Applications

TN + TP
Accuracy = (8)
TN + TP + FP + FN
Sensitivity (Recall) is used to evaluate the model’s ability to predict the True Posi-
tives of each available category.
TP
Sensitivity = (9)
TP + FN
Specificity is used to evaluate the model’s ability to predict True Negatives of each
available category.
TN
Specificity = (10)
TN + FP
Precision is calculated as the ratio between the numbers of True Positive observa-
tions to the total number of Positive observations.
TP
Precision = (11)
TP + FP

4 Fusion of CNN‑LSTM model

The proposed hybrid CNN-LSTM model is shown in Fig. 4. The input parameters are
represented in a range from ­x1 to ­xn. Here n is the count of the overall parameters fed
as input. Data preprocessing is the first step which involves normalizing the data to
eliminate redundancy by data transformation and maintaining compatibility, then filling
up the missed values in the dataset and performing re-sampling that uses samples of the
train set and fit to models with more accuracy. The next stage is optimization. It is also
referred as tuning, which is the procedure of choosing a set of best possible parameters
to be used for learning. Recursive Feature Elimination (RFE) is used for dimensionality
reduction before developing the model. The fundamental process in machine learning or
deep learning pipeline is feature selection which improves the efficiency of the system. The

Fig. 4  Architecture of the developed model based on CNN-LSTM for product return detection
 Multimedia Tools and Applications

RFE uses the filter-based feature selection internally. The algorithm searches for a subset
of features with all the initial features in the training dataset and process by successfully
eliminating the features and resulting in the final subset of desired features. The structure
of this research problem is formulated with three layers which are all fully connected. We
use Rectified Linear Unit activation function (ReLU) on the first two layers. The sigmoid
function is used in the output layer. ReLU activation is used to achieve better performance
compared to sigmoid and Tanh functions. Sigmoid activation is used in the output layer
because the network’s output is between 0 and 1. The architecture of the developed model
based on LSTM is shown in the Fig. 4 with seven inputs (× 1 to xn). The first hidden layer
has twelve nodes and uses the ReLU activation function which is opted as best among the
other activation functions for this proposed model. The Second hidden layer has 8 nodes
and also uses the ReLU activation function. The output layer has 1 node and uses sigmoid
activation function. The Fig. 5 denotes the flow of compiling the model.
The sequential class is used to provide training and the inference features on this
model. We then compile and train the network and specify the loss function using cross
entropy. Optimization is carried out using Adam algorithm. We use Adam optimization
algorithm which is an extension of stochastic gradient descent. Adam is easier to configure,
computationally efficient and produces better results when compared to classical stochastic
gradient descent algorithms.
Figure 6 is the representation of hybrid CNN-LSTM architecture. A CNN-LSTM can
be defined by adding CNN layers on the front end followed by LSTM layers with a dense layer
on the output. The inputs X­ n, n is (0, 1…n) is fed as input and the embedding layer converts
each of the word into a fixed length vector of the defined size. Word2Vec is used to produce
word embedding by learning the meaningful relations and encodes the relatedness into vector
similarity.
The Convolutional layer is used to extort the features from the input which is the major
building block of CNN. The majority of the computations occur in the Convolutional
layer. The output obtained from the convolution operation is termed as feature map,
convolved feature or an activation map. Filter is used as the feature detector and the
feature map is passed through the CNN with all important features. Strides are defined
while creating the CNN which is done by sliding step by step of every element in the

Fig. 5  Flow diagram of Model

compilation
Multimedia Tools and Applications

Fig. 6  Architecture of Hybrid CNN-LSTM model

input data. The channels represented as the sequence of words, as word embeddings and
shape of the words. In general, the text is represented as a array of vectors and processed
using CNN. Here each word mapped to a specific vector in a vector space is composed of
the entire vocabulary. The different kernels applied to the words are merged to a single
vector. We apply ReLU which overcomes the vanishing gradient problem and allows
the developed model to learn faster and perform better and performs better than any other
feature elimination techniques. The max-pooling also termed as down sampling layer, picks the
maximum count of fallings. Pooling ensures that the network detects the features irrespective
of its location and also the size of data passed to CNN is further reduced. The final process in
learning is to use the LSTM model. The LSTM layer is followed by dense layer as the output.
 Multimedia Tools and Applications

The CNN layer is responsible for feature extraction and LSTM for interpreting the features.
LSTM allows learning the word sequences. The dense layer is a deeply connected neural
network layer and is frequently used. It performs classification based on the output from the
LSTM. Before the final output is retrieved the obtained result is passed to Softmax activation
function for the multi-class problems.

5 Results and discussion

The research was proposed by applying CNN-LSTM for building product return classifi-
cation models. The performance of the developed model was observed over e-commerce
consumer reviews dataset. The optimization is performed using the Adam algorithm
which is an extension of stochastic gradient descent algorithms. This Adaptive Move-
ment Estimation Algorithm is a natural successor of AdaGrad and RMSProp which
combines their best properties to form an optimization algorithm [22]. The developed
model is applied to the train and test set. K-Fold cross validation with k = 10 splits is
used to validate the obtained results which is the optimal value for k. The results of the
experiments are illustrated in the Tables 2 and 3 and it is shown that the CNN- LSTM
outperforms in both the cases. The false positives and false negatives rates are reduced
which indicates that the misclassification of non-returns is reduced.
The dataset is split as 80:20 for the training and testing data based on the size on
the dataset which has 34,000 consumer reviews, the CNN-LSTM model has resulted in
the utmost accuracy of 97.67% which is better when compared to the all-other models.
Tables 2 and 3 illustrate the results of the proposed models. Based on the results of eval-
uation metrics as illustrated in Table 2 and we can finalize that the hybrid CNN-LSTM
prototype for product return performs better compared to the other models.
The graphical representation based on the evaluation metrics like precision, specific-
ity, sensitivity, and accuracy is represented in Fig. 7 and it is clearly shown that CNN-
LSTM efficiently handles the high dimensional data with improved accuracy and out-
performs compared to all other models with improved metric score above 90% while
other models like CNN and C-LSTM are ranging below 90%.

Table 2  Product Return Prediction metrics of different models with chosen features
Evaluate Models Accuracy Sensitivity Specificity Precision

Individual Training and Test Set C-LSTM 86.62 83.75 84.41 86.53
CNN 84.73 83.16 82.45 83.17
CNN– LSTM 96.62 97.19 97.37 94.11
K-Fold Cross Validation C-LSTM 90.02 84.21 86.72 88.72
CNN 89.76 87.19 88.94 86.23
CNN – LSTM 97.67 97.23 97.02 97.58
Multimedia Tools and Applications

Table 3  Product Return Prediction metrics of different models including all features
Evaluate Models Accuracy Sensitivity Specificity Precision

Individual Training and Test Set C-LSTM 82.16 82.35 83.19 81.73
CNN 81.75 81.22 80.09 80.03
CNN-LSTM 93.68 92.37 93.02 92.64
K-Fold Cross Validation C-LSTM 86.77 85.42 86.19 86.79
CNN 85.43 86.30 86.71 84.19
CNN-LSTM 97.58 93.16 94.09 93.20

6 Conclusion

The consumer’s sentiment analysis for the purchased products plays a major part
in determining the revenue of the e-commerce firm. CNN learning is used in feature
extraction from the data and usually it uses several Convolutional layers for capturing
the long-term dependencies. The issue arises with the increase of input sequences cap-
turing the dependencies get worsen which leads into deep layers of CNN. The LSTM is
also capable of capturing long-term dependencies. The proposed hybrid CNN-LSTM is
applied for sentiment analysis using the feedback reviews dataset. It results in greater
accuracy of 97.58% when compared to the individual CNN and LSTM models. The

Comparison on Performance Analysis

100

95

90

85

80

75
Accuracy sensitivity speci icity precision

C-LSTM CNN CNN-LSTM

Fig. 7  Performance observation of applied models

 Multimedia Tools and Applications

practical deployment of the model with real-time dataset will result in promising results,
since the model is validated with a maximum accuracy rate.

Author’s contributions Author 1: Vidya Rajasekaran.

He participated in the methodology, Conceptualization, Data collection and writing the study.
Author 2: Dr. Latha Tamilselvan.
He Performed the Analysis the overall concept, writing and editing.

Funding On Behalf of all authors the corresponding author states that they did not receive any funds for this
project.

Data availability All the data is collected from the simulation reports of the software and tools used by the
authors. Authors are working on implementing the same using real world data with appropriate permissions.

Code availability Not Applicable.

Declarations
Ethics approval No ethics approval is required.

Informed consent Not Applicable.

Human and animal ethics Not Applicable.

Conflicts of interest The authors declare that we have no conflict of interest.

References
1. Sherstinsky Alex (2020) Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term
Memory (LSTM) network. Physica D: Nonlinear Phenom 404:132306. https://​doi.​org/​10.​1016/j.​physd.​
2019.​132306. (ISSN 0167-2789)
2. Yong Yu, Si X, Changhua Hu (2019) Jianxun Zhang; A Review of Recurrent Neural Networks: LSTM
Cells and Network Architectures. Neural Comput 31(7):1235–1270. https://​doi.​org/​10.​1162/​neco_a_​
01199
3. TejaKattenborn Jens Leitloff, Schiefer Felix, Hinz Stefan (2021) Review on Convolutional Neural Net-
works (CNN) in vegetation remote sensing. ISPRS J Photogramm Remote Sens 173:24–49. https://​doi.​
org/​10.​1016/j.​isprs​jprs.​2020.​12.​010
4. Yu Shujuan et al (2021) DPTCN: A novel deep CNN Model for short text classification. IFS
41:7093–7100
5. Fesseha Awet, ShengwuXiong Eshete D, Emiru Moussa Diallo, Dahou Abdelghani (2021) Text Clas-
sification Based on Convolutional Neural Networks and Word Embedding for Low-Resource Lan-
guages: Tigrinya. Information 12(2):52. https://​doi.​org/​10.​3390/​info1​20200​52
6. Rehman AU, Malik AK, Raza B, Ali W (2019) A Hybrid CNN-LSTM Model for Improving
Accuracy of Movie Reviews Sentiment Analysis. Multimed Tools Appl. https://​doi.​org/​10.​1007/​
s11042-​019-​07788-7
7. Chen R, Wang X, Zhang W, Zhu X, Li A, Yang C (2019) A hybrid CNN-LSTM model for typhoon
formation forecasting. GeoInformatica. https://​doi.​org/​10.​1007/​s10707-​019-​00355-0
8. Alhussein M, Aurangzeb K, Haider SI (2020) Hybrid CNN-LSTM Model for Short-Term Individual
Household Load Forecasting. IEEE Access 8:180544–180557. https://​doi.​org/​10.​1109/​access.​2020.​
30282​81
Multimedia Tools and Applications

9. She X, Zhang D (2018) Text Classification Based on Hybrid CNN-LSTM Hybrid Model, 2018 11th
International Symposium on Computational Intelligence and Design (ISCID), pp. 185–189. https://​doi.​
org/​10.​1109/​ISCID.​2018.​10144
10. Tasdelen A, Sen B (2021) A hybrid CNN-LSTM model for pre-miRNA classification. Sci Rep
11:14125. https://​doi.​org/​10.​1038/​s41598-​021-​93656-0
11. Kumar KLS, Desai J, Majumdar J (2016) Opinion mining and sentiment analysis on online customer
review. IEEE International Conference on Computational Intelligence and Computing Research
(ICCIC) 2016:1–4. https://​doi.​org/​10.​1109/​ICCIC.​2016.​79195​84
12. Rahardja U, Hariguna T, Baihaqi WM (2019) Opinion Mining on E-Commerce Data Using Sentiment
Analysis and K-Medoid Clustering, 2019 Twelfth International Conference on Ubi-Media Computing
(Ubi-Media), pp. 168–170. https://​doi.​org/​10.​1109/​Ubi-​Media.​2019.​00040
13. Basani, Sibuea HV, Ida PatonaSianipar S, PreslySamosir J (2019) Application of Sentiment Analysis
on Product Review E-Commerce. Journal of Physics: Conference Series 1175:012103. https://​doi.​org/​
10.​1088/​1742-​6596/​1175/1/​012103
14. Sari PK, Alamsyah A, Wibowo S (2018) Measuring e-Commerce service quality from online customer
review using sentiment analysis. J Phys: Conf Ser 971:012053. https://​doi.​org/​10.​1088/​1742-​6596/​
971/1/​012053
15. Ligthart A, Catal C, Tekinerdogan B (2021) Systematic reviews in sentiment analysis: a tertiary study.
ArtifIntell Rev 54:4997–5053. https://​doi.​org/​10.​1007/​s10462-​021-​09973-3
16. Sarker IH (2021) Machine Learning: Algorithms, Real-World Applications and Research Directions.
SN Comput Sci 2:160. https://​doi.​org/​10.​1007/​s42979-​021-​00592-x
17. Mohammed B (2020) Natural Language Processing Applications in Business In E-Business: Higher
Education and Intelligence Applications, edited by Robert Wu, MarinelaMircea. London: IntechOpen.
https://​doi.​org/​10.​5772/​intec​hopen.​92203
18. Lin X (2020) Sentiment Analysis of E-commerce Customer Reviews Based on Natural Language Pro-
cessing. Proceedings of the 2020 2nd International Conference on Big Data and Artificial Intelligence.
https://​doi.​org/​10.​1145/​34362​86.​34362​93
19. Zhang X, Chen F, Huang R (2018) A Combination of RNN and CNN for Attention-based Relation
Classification. Procedia Computer Science 131:911–917. https://​doi.​org/​10.​1016/j.​procs.​2018.​04.​221.
(ISSN 1877-0509)
20. Alzubaidi L, Zhang J, Humaidi AJ, et al (2021) Review of deep learning: concepts, CNN archi-
tectures, challenges, applications, future directions. J Big Data 8:53. https://​doi.​org/​10.​1186/​
s40537-​021-​00444-8
21. Wang W, Gang J (2018) Application of Convolutional Neural Network in Natural Language Process-
ing. International Conference on Information Systems and Computer Aided Education (ICISCAE)
2018:64–70. https://​doi.​org/​10.​1109/​ICISC​AE.​2018.​86669​28
22. Ahmad I, Zheng W, Luo J, Liou M (2006) A fast adaptive motion estimation algorithm. IEEE Trans
Circuits Syst Video Technol 16(3): 420-438. https://​doi.​org/​10.​1109/​TCSVT.​2006.​870022

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.