Journal of Ambient Intelligence and Humanized Computing

https://doi.org/10.1007/s12652-020-02234-1

ORIGINAL RESEARCH

Optimized machine learning based collaborative filtering (OMLCF)

recommendation system in e‑commerce
J. Anitha1 · M. Kalaiarasu1

Received: 24 March 2020 / Accepted: 17 June 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
A recommender system (RS) is a subcategory of an information filtering system that attempts the prediction of the score or
the importance given to an item by a user. RS has garnered the attention of the business community and individuals towards
itself owing to its significance in the e-commerce field. One of the most common methods of the RS used for the genera-
tion of recommendations is the CF technique (collaborative filtering). But, CF-based RS yields untrustworthy similarity
information and yields a recommendation quality that is not satisfactory. Support vector machine (SVM) helps in enhancing
issues in the CF technique. The parameter of the SVM algorithm minimizes the system’s accuracy, and therefore in classifier
improved ant colony optimization (IACO) is brought-in for parameter optimization. In the newly introduced system, RS will
be carried out in two stages which include (1) SVM classifier for classifying the entities into positive and negative feedback.
The best value achieved indicates the optimized values of the parameters of SVM employing the IACO algorithm, which
are given in the form of an input to the classifier to carry out pair-wise classification, (2) then, we construct SVM–IACO
based collaborative filtering algorithm. The collaborative filtering recommendation’s execution is only done on the entities’
positive-feedback. The actual content used for recommendation is highly reduced owing to the classification much earlier;
therefore the collaborative filtering improves the efficiency in comparison with the classical one. Tests on Taobao data (an
Alibaba owned Chinese online shopping website) revealed that the algorithm yields a superior recommendation accuracy
thereby commanding a particular predominant place in the e-commerce field.

Keywords  Collaborative filtering (CF) · Support vector machine (SVM) · Improved ant colony optimization (IACO) ·
E-commerce · Recommender system (RS)

1 Introduction that e-commerce websites start finding a problem solution

of "information overload" using recommender systems (RS)
The exponential rise in the distributed information’s count (Zhao 2019).
over the World Wide Web has rendered it hard searching Commodity recommender systems can be used for
and finding objects on user-interested things. The fast-paced recording the representative and behavioral information of
development of the Internet and its impact on mobile has the user like buying and browsing. Through the analysis of
made the e-commerce industry thriving on the widespread the information acquired from the models pertaining on the
interest from every walk of life. The magnanimous amount user’s choice, commodity resources, which suit the user’s
of information in the e-commerce field and the extraction strong demand or may filter the provoked curiosity in them
of information that are very much helpful has led to the from the commodity information present on web sites which
emergence of recommender systems since it is now expected have e-commerce, and later made as recommendations to
users (Lin and Wenzheng 2015).
* J. Anitha RS has garnered the attention of the business community
[email protected] and individuals towards itself owing to its criticality and sig-
* M. Kalaiarasu nificance in the e-commerce field and gain on the acknowl-
[email protected] edgment of a potential customer. RS provides basic services,
which are the empirical solutions on the recommendations
1
Department of Information Technology, Sri Ramakrishna made for different application structures and services. In the
Engineering College, Coimbatore, India

13
Vol.:(0123456789)
 J. Anitha, M. Kalaiarasu

cutting edge time of propelling web advancements, the RS the optimized values of the SVM parameters employing the
has turned the notification of the business society and the IACO algorithm, which are submitted as an input to the clas-
regular man towards itself because of its essentialness and sifier for the pair-wise classification process. (2) In the next
significance in the E-trade (e-commerce) and accomplish- step, SVM–IACO based collaborative filtering algorithm
ment of prevalent client’s endorsement (Yan 2017). is built. Therefore, the proposed approach provides better
In recent times, e-commerce is found to have a strong results for the recommendation system.
connection to the satisfaction felt by the customer, and suc-
cess is always based on customer trust ultimately. RS tries
to service the users with the recommendations on products,
services, and information as per their interest, consider- 2 Literature review
ing their requirements or choices. Recently, RS has gained
increasingly widespread attention and has been used in dif- Li and Li (2019) studied a modified collaborative filtering
ferent fields (Prando et al. 2017). algorithm for meaningful exploration and coordinated rec-
A recommender system is an information filtering tech- ommendation technology. This research work is performed
nology utilized for providing helpful recommendations to by following three important steps, (1) first, a modified col-
a set of users for items or services that they might be inter- laborative filtering algorithm, (2) in the second step, we
ested in. The recommender framework utilizes a sifting algo- examine the algorithm of community detection, and algo-
rithm on the info appraisals and gives recommendations on rithm of two overlapping community detection that depends
the new things (i.e., target things) for the dynamic client on the node in the central and k-based faction, which helps
(active). Collaborative filtering (CF) is one of the notable in effective community mining in the network, (3) finally, a
methodologies in recommender frameworks (Li et al. 2016). segment of client networks chose from the client arrange is
CF refers to the procedure of item filtering for exploiting sent by the client thing system to be the applicant neighbor-
the choices that are known of a user’s settings. These CF ing client bunch for the objective client, thus minimizing the
algorithms can fall into two categories, i.e., item-based and computation time and improving the speed recommendation
user-based recommender systems. The former one performs and accuracy of the recommendation system.
the comparison of similarities between items, and the latter This research work perfectly integrates the social network
one rather provides the comparison of the user similarities technologies and technology of collaborative filtering, which
in the process of recommendation (Wasid and Kant 2015). can help in a huge improvement in the recommendation
Conventional collaborative filtering is marred by the draw- system’s performance. The dataset of Movie Lens is used
backs of recommending just based on a single indicator, for testing two performance indexes which consist of mean
which are either the scores or comments of the user. But, absolute error (MAE) and root mean square error (RMSE).
the scores and comments must not be analyzed individually, The experimental result reveals that the modified collabo-
since the inconsistency between the scores and comments rative filtering algorithm offers much better performance
are frequently observed practically (Ramzan et al. 2019). compared to other algorithms of two collaborative recom-
Hence, classical collaborative filtering is deficient in terms mendation in terms of performance of MAE and RMSE.
of the predictive accuracy metric. Zarzour et al. (2018) studied the k-means-singular value
To find a resolution to these issues, a new algorithm that decomposition (k-means-SVD) based on the algorithm of
merges filtering in a collaborative way and support vec- designed collaborative filtering recommendation for dimen-
tor machine (SVM) is utilized in the available techniques. sionality reduction and clustering approaches.
Diverse from classical collaborative filtering, SVM is uti- In the first step, the algorithm of k-means clustering was
lized for the classification of products into positive and used for clustering the users in the identical group as per
negative feedback. Thus, SVM and recommendation of col- the choices, and later the SVD was utilized in every cluster
laborative filtering are integrated to increase the competence not just as an approach of dimensionality reduction but also
and correctness associated with the recommendation. But, in the form of a potential technique, which could be hugely
CF-based RS yields untrustworthy similarity information, useful in getting the similar users. It introduces and assesses
and its recommendation quality is not quite up to the mark. a resourceful two practical datasets for movie recommen-
For resolving this problem of the parameter of the SVM dation known as Movie Lens 1 M and Movie Lens 10 M,
algorithm reducing the accuracy of the system, improved which can produce recommendations that are accurate and
ant colony optimization (IACO) is introduced for the opti- hugely superior. RMSE metric was used for computing the
mization of parameters in the classifier (Che et al. 2019). prognostic accuracy of the k-means-SVD technique. The
RS will be carried out in two steps, which include (1) SVM results of experiments reveal that this novel technique helps
classifier for the classification of commodities into positive in considerably improving the performance achieved in the
and negative feedback. The best value so acquired indicates recommendation systems.

13
Optimized machine learning based collaborative filtering (OMLCF) recommendation system…

Wang et al. (2018) suggested a customized recommen- drawbacks of one single algorithm and boost up the concert
dation algorithm that depends on the feedback of User’s in terms of both quality recommendation and effectiveness.
Implicit (abbreviated as BUIF). BUIF not just considers the Feng et al. (2018) presented a modified similarity model
trends in the user’s purchase but also the behavior of com- for collaborative filtering, which considers three significant
parison and commodity-sequences. The behaviors of pur- factors of similarity for reducing the difference in similarity
chase and comparison and commodity-sequences from the computation. In comparison with the conventional similarity
behaviors of users log are extracted; the similarity of the user measure, the benefits of the modified similarity model are
is calculated using the behaviors purchase and comparison that it exploits the rating data and resolves the issues involv-
and word-embedding is extended to the embedding of items ing co-rated commodities.
to get the commodities similarity. This research work renders recommendations to the users
Few algorithms in the natural language processing (NLP) via three stages: at first, the ratings produced by the nature
field, like term frequency–inverse document frequency of the interactions of a user are obtained and then stored in
(TF–IDF) and word2vec were modified and used for cal- the database. Afterward, the technique of k-Nearest Neigh-
culating the similarities of user’s and item’s, which renders bors (kNN) is used for score prediction of the target user’s
the user’s and item’s similarity to achieve more accuracy. unscored commodities.
Besides, a secondary reordering screening process was The drawback of kNN is the means of calculating the
built for getting the recommended commodities finally. similarities between the target user and his/her neighbors. A
The results metrics include precision, recall, and F-score modified model of similarity will reduce the similarity vari-
for evaluating the quality obtained in the algorithms. The ations computation and enhances the recommendation accu-
experimental results on the dataset of JData reveal that algo- racy, which will be presented as follows. Experiments were
rithm BUIF exhibits a better increase in terms of accuracy carried out on four datasets, for confirming the enhanced
when distinguished with other CF algorithms. similarity model’s effectiveness. The results reveal that the
Hwangbo et al. (2018) shown a K-RecSys framework modified similarity model can help in improving the choices
which improves the conventional item-specific collabora- of the recommender system efficiently and it is apt for the
tive filtering calculation. K-RecSys incorporates the online sparsity data.
item click information and disconnected item deal informa- Jiang et al. (2019) introduced an incline one algorithm
tion weighed to delineate on the web and disconnected the that relies upon the mix of confided in information and cli-
selection of purchasers. It additionally utilizes an inclination ent comparability, which can be actualized in various rec-
rot work for indicating the varieties in decisions over the ommender frameworks. This calculation comprises of three
progression of time, and finally, gives suggestions to sub- stages. Right off the bat, the slope one algorithm that relies
stitute and correlative items utilizing the data of item type. upon dependable information is contrasted and the custom-
Furthermore, an A/B test is directed in the genuine ary algorithm.
operational condition for contrasting the K-RecSys and the Also, the contrast between the slope one algorithm that
accessible cooperative sifting framework acknowledged with relies upon the blend of solid information and likeness and
online information as it were. The aftereffects of test dis- the algorithm that relies upon believed information utilizing
close that the K-RecSys framework displays prevalent execu- MAE is dissected. In the third step, the slope one algorithm
tion as far as item snaps and deals in the web-based shopping relying upon the mix of solid information and similitude and
center and its suggestions for a substitute are utilized all the the algorithm relying upon dependable information utiliz-
more frequently contrasted with integral proposals. ing RMSE is looked at. This comparability is added to the
Lu et al. (2015) presented an algorithm for the hybrid weight factor of the altered slope one algorithm, and further,
recommendation that depends on modified collaborative fil- the last proposal condition is gained. Finally, the correlations
tering of clustering the user context fuzzy and content-based. between the novels incline slope one algorithm relying upon
In the case of collaborative filtering, first, the user classifica- client likeness under various measured datasets.
tion is performed based on fuzzy clustering by the context Hu et al. (2019) exhibited Collaborative Filtering Based
of the user, and thereafter collaborative filtering is utilized Recommendation Algorithms (CFRAs); content-based pro-
for the recommendation of products for users with similarity. posal calculations; and half and half suggestion calculations.
Also, the modified content-based algorithm establishes The community-oriented filtering algorithm is remembered
the user’s vector feature and items dynamically. For exam- for the general proposal methods for internet apparel. The
ining the performance of the hybrid recommendation test dataset is worked with a viable enormous scope occur-
technique, an experiment is conducted on the Movie Lens rence from one of the greatest business-to-shopper online
Dataset. The evaluation metrics include precision/recall and business stages working in China. Taking the buying events
coverage is used in the results of the experiment. They dis- and the converse client recurrence, three closeness measures
close that the algorithm for hybrid things can prevent the are intended for the cosine-based comparability algorithm.

13
 J. Anitha, M. Kalaiarasu

A few numerical examinations are done to assess the sug-

It is not quite an active learner, implying that it does not use the

Ignore the social relationships. Easily cold start problem more

gestion procedures and their exhibitions are assessed as far

training data for learning anything and just uses the training
are perceived based on assumption, so it may not be widely
It has two major drawbacks that are sparsity, and scalability.

Quite expensive and it is not always accurate, so The results

The drawback of content based filtering is the ability b to
differentiate between a bad and a good item based on the
as three criteria, for example, precision, recall, diversity. The
decrease of purchaser merchandise co-event lattice helps to
improve the computational exhibition by considering the
most indistinguishable clients because the test occasion is
enormously scaled.
Utilizing this commonsense model, the calculations
are inspected under the assessment conditions. The conse-

It totally depends on human ratings

data for classification purposes
quences of tests show that the suggestion dependent on cli-

information that is retrieved

ent recurrence comparability is very alluring for the internet

Sparsity (insufficient data)

dress proposal; the co-event framework decrease technique
is effective for improving the presentation of the proposal.
For approving the CFRAs methods, the correlation is done

(low accuracy)
Disadvantages
on the Movie Lens dataset.

Scalability
accepted
Chang et al. (2019) proposed a new algorithm that inte-
grates filtering collaboration filtering and support vector
machine (SVM). In contrast with filtering in classical col-
laboration, SVM is used for the commodities classification

Improving the choices of the recommender system efficiently

into positive and negative feedback. After this, the com-

Prevent the drawbacks of one single algorithm and boost up

the concert in terms of both quality recommendation and
Considerably improving the performance achieved in the
modities, which exhibit positive feedback is chosen to com-
pute the elaborate scores of marks and comments. Then, we
develop the algorithm for SVM-based collaborative filtering.
Experiments are carried out on Taobao data (an Alibaba
owned Chinese online shopping website) revealed that the

Provides better results of MSE, recall etc.

algorithm offers good recommendation precision and effec-
tiveness, thereby commanding a specific practical role in

and it is apt for the sparsity data

the e-commerce field. Table 1 shows the comparison of the
recommendation systems

existing approaches.
The drawback of the k-NN algorithm is that it is not quite
Advantages and results

an active learner, implying that it does not use the training

data for learning anything and just uses the training data for
effectiveness

classification purposes.
K-means clustering algorithm is (1) prediction of k-value
is hard. (2) It does not yield good performance with global
clusters. (3) It does not function well with clusters (in the
actual data) of various sizes and density. To resolve these
Collaborative Filtering Based Recommendation Algorithms
Hybrid recommendation that depends on modified collabo-

problems, the support vector machine (SVM) based clas-

Modified similarity model for collaborative filtering with
rative filtering of clustering the user context fuzzy and

sifier is introduced for collaborative filtering classification.

Table 1  Comparison of the existing approaches

3 Proposed methodology
K-means-singular value decomposition

The system RS proposed will be carried out in two steps (1)

SVM classifier for classifying the commodities into positive
and negative feedback.
The issues involving low efficacy in the SVM classi-
fier requires improvement. The best value so acquired
K-NN algorithm

indicates the optimized values of the SVM parameters

content-based

employing the IACO algorithm, (2) then, SVM-IACO

Approaches

(CFRAs)

based collaborative filtering algorithm is built. Python-

based Scrapy experiments were used for extracting data
of online commodity information on the Taobao dataset.

13
Optimized machine learning based collaborative filtering (OMLCF) recommendation system…

Fig. 1  Architecture of e-com- Client Client

merce recommendation system

Request Response Request Response

Web Server

Data format conversion

Strategy algorith m User interest

Configurat ion Modeling

User need Online

analysis recommendation

Reco mmended User interest

strategy library model library

Reco mmendation engine

13
 J. Anitha, M. Kalaiarasu

Fig. 2  Proposed framework of
e-commerce recommendation
Taobao database which includes
system
7 categories of commodities

E-commerce Commodities

Recommender System (RS)

SVM Classifier based Collaborative filtering

Training set Testing set

Divide commodities into positive-feedback commodities and negative- Filtering out negative feedback and reserving
feedback commodities positive feedback

To optimize parameters of SVM classifier using IACO

Good recommendation accuracy and

recommendation efficiency

Figure 1 illustrates the proposed framework of the e-com- problem of classification can be considered to be an optimi-
merce recommendation system. zation problem where the objective is to get the best model
Figure 2 shows the proposed framework of e-commerce representing the predictive associations in the data. Sup-
recommendation system. port vector machine (SVM) has achieved more focus and
has been used in the problems involving data classification
3.1 Support vector machine (SVM) based for finding a better solution. The selection quality of SVM
collaborative filtering parameters and kernel functions influences the performance
of learning and generalization (Syarif et al. 2016).
3.1.1 Commodity information acquisition
3.1.2.1  Procedure of  SVM The mapping of the sam-
For verifying the real impact of modified collaborative ple’s training from the input space onto a higher dimen-
filtering in the e-commerce field, Python-based Scrapy is sional feature space through a mapping function ϕ
used for acquiring the online commodity information on was the
( main ) idea of SVM. Provided a training set
Taobao, which primarily includes the commodity name, S = { xi , yi |xi ∈ H, yi ∈ {±H}1}, i = 1, 2, ..., l} , where xi
commodity information and the comments of the user on refers to the vector’s input and yi indicates labels of the xi ,
it. Also, based on a large comment’s count, translates Tao- the target function is
bao into the definitive value of commodities using analysis
of semantics, like superior stuff, quick logistics, etc. ⎧ ∑l
⎪ min 𝜙(w) = 21 w ⋅ w + C 𝜉i
⎨ � � � � i=1 (1)
3.1.2 Support vector machines (SVM) ⎪ s.t.yi w ⋅ 𝜙 xi + b ≥ 1 − 𝜉i , 𝜉i ≥ 0 i = 1, 2, … , l,
⎩
In the supervised learning approach, we have classification where C stands for a penalty parameter, 𝜉i refers to
which learns a function from the training data set com- non-negative slack variables. The decision function can
prising of input features/attributes and group output. The be given as:

13
Optimized machine learning based collaborative filtering (OMLCF) recommendation system…

[ l ]
∑ ( ) categorizes the commodities into two groups. Thus, the data
f (x) = sign yi 𝛼i K xi ⋅ x + b (2) depicting the dislike of the users are removed and just the data
i=1 indicating the users’ likes are stored.
The typical kernel functions utilized in SVM are given 3.1.2.3  Parameters of  SVM for  optimization The SVM
as below: execution shows the capability of sorting the obscure
Linear kernel information tests directly through SVM characterized by
( ) gaining from the training samples otherwise called spec-
K x, xi = x ⋅ xi (3) ulation ability. Regularization consistent C and kernel
Polynomial kernel function parameters substantially affect SVM’s specula-
tion capacity. The regularization consistent C chooses the
( ) ( )d
K x, xi = x ⋅ xi + c (4) tradeoff between the decrease of fitting blunder and the
expansion of order level. The kernel function parameters,
RBF kernel: similar to the transmission capacity σ(sigma) of the RBF
( ) kernel, apply an effect on the mapping changes of informa-
( ) x − xi2 tion space and adjusts the intricacy level of test circulation
K x, xi = exp − (5)
2𝜎 2 in the higher dimensional component space. Since the esti-
mation of each parameter, excessively high or excessively
less, all affect the SVM generalization, the optimization of
3.1.2.2  Classification  The set of data used in the experi- parameters is vital and advantageous for attaining a good
ment primarily consists of marks and values of represent- generalization capability practically. This research work
ative commodities. A 2500-dimension vector set is con- introduces an IACO algorithm for the automatic optimiza-
structed as per the data, and the representative value which tion of the parameters C and k (Zhang et al. 2010).
is missing and zero is filled there.{(Assume ) that the set} Drawbacks of SVM:
of data commodities on Taobao is xi , yi |i = 1, 2, … n ;
and the available { for recommendations
} ( of a set of )data • The important drawback of SVM is the kernel choice.
commodities is xi |i = 1, 2, … n ;xj = xj1 , xj2 , … , xjk and The incorrect selection of the kernel can result in aris-
xi = (xi1 , xi2 , … , xik ) indicates the representative attrib- ing in the error degree.
utes of set i and set j; yi { ∈ − 1, 1} stands for the type of • When the number of samples is higher, it yields limited
output. yj = −1 Implies that the commodity has adverse performance.
feedback; yj = 1 implies the commodity has feedback with • SVMs offer good generalization performance, however,
positive comments. SVM constructs {( the )model for clas-} it can be dead slow in the test stage.
sification with data set commodity xj , yj |j = 1, 2, … , n
to get the optimal hyperplane g (x) = w ⋅ x + b = 0. To resolve these problems, the parameters of SVM are
In this experiment of respective optimization of SVM optimized using the improved ant colony optimization
is computed as: (IACO) algorithm (Jaganathan et al. 2011).

∑
n
1∑
n
( ) n
∑
max L(w, b, 𝛼) = 𝛼j − 𝛼j 𝛼q yj yq K xj xq s.t. yj 𝛼j = 0; 0 ≤ 𝛼j ≤ C (6)
𝛼
j=1
2 j=1 j=1

k(.) refers to the radial basis function; optimization 3.2 Support vector machine–improved ant colony
∑
n
optimization (SVM–IACO) based collaborative
helps to get the optimal solution𝛼j∗ . w∗ = 𝛼j∗ yj xj . Hence,
j=1 filtering algorithm
we can estimate the classification decision function of
optimal hyperplane definition through: 3.2.1 Optimization
( n )
∑ ( )
f (x) = sgn(g(x)) = sgn ∗
𝛼j yj k xj , x + b ∗
(7) Swarm intelligence is a computational and behavioral
j=1 metaphor for problem-solving that takes inspiration from
the social behavior of insects or other animals. ACO is
Two groups of commodities were available. When f (x) = −1 , one of the most powerful optimization methods that take
the commodity has adverse feedback; whenf (x) = +1 , positive inspiration from the foraging behavior of real ants.
feedback is provided for this commodity. SVM classification

13
 J. Anitha, M. Kalaiarasu

Swarm insight is a computational and conduct simili- and heuristic capacity is, the higher the likelihood of the
tude for critical thinking that takes motivation from the parameters of SVM to be picked is.On the other hand, the
social conduct of creepy crawlies or different creatures. classification accuracy on the map will hoard and disappear
ACO is one of the most impressive advancement tech- over time and the classification accuracy update formula is
niques that take motivation from the searching conduct expressed by
of genuine ants.
𝜏ij (t + Δt) = (1 − 𝜌)𝜏ij (t) + Δ𝜏ij (t), (9)

3.2.2 Ant colony optimization (ACO) ∑

m
Δ𝜏ij (t) = Δ𝜏ijk (t) (10)
In nature, ants store characterization exactness on the k=1

ground to convey one another. The kept order classifica- where ρ is the order exactness dissipation coefficient tak-
tion accuracy encourages the ants to locate the briefest ing qualities over the interim [0,1] to stay away from an
way between the parameters of SVM and the enhanced excess of a gathering of characterization classification accu-
qualities. All the more explicitly, in looking for improved racy; Δτij (t) represents the steady arrangement exactness on
qualities, ants smell the grouping exactness left by past the way from i to j after the time t. The underlying time of
ants of a similar settlement and will, in general, follow the the Δτij (0) is 0. Δτkij (t) denotes the addition of characteriza-
ways set apart by solid arrangement classification accuracy
tion classification accuracy left by subterranean insect k after
focuses. As such, ants pick their way by a probabilistic
t time on the way from i to j, which is characterized as
choice guided by the measure of arrangement classification
{
accuracy: the bigger the measure of characterization exact- Q
if (i, j) ∈ tour done by the ant k,
(11)
k
ness on a path, the higher the likelihood that ants follow Δ𝜏ij (t) = Lk
0
that trail while picking their way. As the expansion of the
ants strolling through a similar way, the order exactness where Lk means the length of the sum of the path con-
fixation on the way will increment. Right now, the likeli- sidered by the ant in this cycle. Q stands for the intensity
hood that ensuing ants pick such a way will increment. classification accuracy, which affects the convergence speed
Furthermore, ants would locate another way as indicated of the algorithm.
by the difference in the earth (Gohari et al. 2017).With Drawbacks of ACO:
the point of improving the way arranging proficiency, the
idea of the heuristic capacity η and tabu rundown tabuk 1. The system exhibits poor initialization and slower con-
were introduced into the artificial ant colony model. All vergence rateand consumes a huge amount of time.
the SVM parameters of the insect went through would be 2. Owing to the deficit of correct guidance in the early
spared in the tabu rundown to ensure that the ants won’t phase of the population generation, much longer time
return and encompass in a similar area. The ant’s search to get multiple efficient paths.
parameters of SVM close by during the time spent devel-
opment dependent on the changing likelihood. The state The heuristic function and tabu list are presented in the
progress rule, otherwise called foraging rule, is communi- improved ant colony optimization (IACO) for resolving
cated as underneath these problems to a specific level.
⎧ � �𝛼 �
𝜏ij (t) ⋅ 𝜂ij (t)
�𝛽
⎪ ∑ if j ∈ allowedk
(8)
k
𝜌ij = ⎨ [𝜏is (t)]𝛼 ⋅[𝜂is (t)]𝛽 ,
⎪
s∈allowedk

0 else 4 improved ant colony optimization (ACO)

⎩
The improved ACO completes its pursuit dependent on
where 𝜏ij (t) represents the order classification accuracy the changing likelihood (1). In any case, contrasted with
focus on the way i to j, s is the present parameters of SVM, the accessible algorithms, the essential distinction is that
𝜂ij = 1∕dij the proportional of the separation among i and when there are just boundary matrices or parameters of
j. An alleged tabu rundown tabuk is the arrangement of SVM in tabu table close by, the ant is permitted to come
parameters of SVM that are picked by ants. Permitted is back to the past parameters of SVM, which is known as the
the arrangement of parameters of SVM that ants can pick rollback methodology. Under the rollback methodology,
in the subsequent stage. α and β speak to the impact of if the past parameters of SVM are in the tabu rundown,
arrangement classification accuracy concentration and the there won’t be any open parameters of SVM for the ants.
heuristic information, separately. The above recipe shows At that point, the ants will be named as the invalid ants
that the more prominent the grouping classification accuracy

13
Optimized machine learning based collaborative filtering (OMLCF) recommendation system…

{
and the left classification accuracy won’t be refreshed, 0 if 𝜏(i,j) = 0
which offers to ascend to the passing procedure (Li et al. fit(i, j) = i = 1, 2, … n; j = 1, 2, … n. (14)
1 else
2014).A noticeable bit of leeway of the rollback method-
ology is that such a technique can successfully amplify where (i, j) is the framework. (i, j) is the convergence of
the measure of the ants to the objective point, and in this arrangement exactness. The connection of the considerable
manner improves the convergence of compelling group- number of parameters of SVM that exist in order classifica-
ing exactness. By using the passing methodology (death tion accuracy is the way that is wanted.
strategy), the deceptive of arrangement exactness to the Step 1: System initialization. Set the quantity of every
subterranean insect state can be decreased. In ACO, the age of ants be m, the complete wave number be n, the
refreshing of arrangement classification accuracy relies most extreme number of emphases being, the arrangement
mostly upon the number of ants going through the way (i,j) classification accuracy weight be α, the heuristic capac-
at the minute t. In any case, some of them are invalid ants. ity weight be β, the underlying order exactness be 0, the
At the end of the day, the ants don’t arrive at the objective characterization exactness vanishing coefficient be ρ, the
point even though they have ever gone through this way. grouping classification accuracy force be Q, individually,
Accordingly, the undifferentiated characterization classi- and include the beginning stage of S to the tabu rundown.
fication accuracy left by these insufficient ants will mis- Step 2: Select the following parameters of SVM j as
direct the advancement of the ant colony (Rehman et al. indicated by the transition probability (1) and spot the
2018). The reasons talked about above vigorously weaken chose parameters of SVM in the tabu rundown. If the goal
the exhibition of the algorithm on the advancement time point is set in the tabu rundown, bounce to Step 4.
of the subterranean insect state just as the union rate (Wu Step 3: The ant comes back to the previous parameters
et al. 2018).Another order exactness refreshing standard of SVM; if there are no elective parameters of SVM close
is executed for countering the symptoms on the algorithm, by, at that point put the first parameters of SVM into the
which alludes to the Max–Min ant framework. After one tabu rundown; Step 2 is re-executed. The order classification
cycle is done, just the order classification accuracy left by accuracy on the way is subtracted from the fixed worth and
the ants arrive at the objective point is included will be the following emphasis is executed, on the off chance that
refreshed. In this way, the way length assumes a focal job the parameters of SVM are as yet unselected.
for the ACO. The new refreshing guideline is given by. Step 4: Update classification accuracy.
( ) Step 5: If n ≥ ng , stop and save the final output path,
t otherwise jump to Step 2.
𝜏ij
+1
= (1 − 𝜌)𝜏ij (t) + Δ𝜏ijbest (t), same (12)

Δ𝜏ijbest (t) =
Q 5 Results and discussion
(13)
Lbest
5.1 Data preparation
where Δτbest
ij
(t) represents the augmentation of the grouping
exactness left by the insect which has discovered the ideal For the information extraction of online data of ware on Tao-
way, Lbest is the complete length of the way from the begin- bao, the investigation utilizes Python-based Scrapy which
ning stage to the objective point, Q represents the arrange- includes 7 gatherings of wares (garments, books, appara-
ment classification accuracy power. Rather than refreshing tuses, advanced items, cell phones, shoes, and packs) and
the grouping classification accuracy at each search, the almost 34,000 bits of broad remarks. There are 4000 lumps
improved algorithm refreshes the characterization exactness of information for every set, among which 2500 pieces are
after every cycle. viewed as the training set and the remaining are used for
Improved ACO propelled by the MAX–MIN ant frame- testing. The experimental results are implemented in MAT-
work, setting the huge beginning estimation of the cen- LAB 7.0.
tralization of characterization classification accuracy. The
grouping exactness focus will diminish with the number 5.2 Assessment indicator
of cycles and the order classification accuracy conver-
gence of the way that no ant passes will be decreased to 0. Predictive accuracy P represents the probability that the
At long last, the algorithm yields all current arrangement user may like an item in the recommendation list, which can
classification accuracy for parameters of SVM, which is show the accuracy of the recommender system. The formula
the last way. The target capacity of the advancement issue to calculate predictive accuracy of recommender system is
is as as follow:

13
 J. Anitha, M. Kalaiarasu

Table 2  Predictive accuracy P Methods N (number of different recommended items)

of recommender system with
diverse N values 5 10 15 20 25 30

K-RecSys–CF 33 26 23 22 21 15
SVM–CF 36 30 26 24 22 18
SVM–IACO–CF 39 32 28 26 24 20

Table 3  Recall Rate R of the Methods N (number of different recommended items)

recommender system with
different N values 5 10 15 20 25 30

K-RecSys–CF 13 19 24 28 33 35
SVM–CF 16 24 26 31 35 37
SVM–IACO–CF 18 26 28 33 37 39

Table 4  F-measure F of Methods N (number of different recommended items)

recommender system with
different N values 5 10 15 20 25 30

K-RecSys–CF 17 20 22 23 24 25
SVM–CF 21 25 26 27 28 30
SVM–IACO–CF 23 27 28 29 30 32

1∑ 1 ∑ ||RLu ∩ TLu ||
m m 45
P= Pu = (15)
m u=1 m u=1 N 40

Recall rate R indicates the ratio of items that are liked by 35

Predictive accuracy P(%)

K-RecSys-CF
the users in the list of recommendations, which can reflect 30
SVM-CF
the degree of satisfaction of the users with the outcome’s
25
recommendation. The higher would be the user’s satisfaction
SVM-IACO-CF

degree, the greater the recall rate. The expression to compute 20

the recall rate of the recommender system is as given: 15

1∑ 1 ∑ ||RLu ∩ TLu ||
m m
10
R= Ru = (16)
m u=1 m u=1 TLu
5

F–measure is utilized for assessing the recommendation 0

performance of the algorithm on the whole. The stronger 5 10 15 20 25 30

recommending capability of the algorithm, for higher Recommended item number N

F-measure entails. The formula to compute F-measure of

the recommender system is as given: Fig. 3  Predictive accuracy P of three models with different N values

1∑ 1 ∑ 2 ∗ Pu ∗ Ru
m m
F= Fu = (17) SVM–CF correspondingly with different values of rec-
m u=1 m u=1 Pu + Ru ommended items (N) is different values are tabulated in
Tables 2, 3, and 4.
Figure 3 illustrates the performance outcomes of predic-
5.3 Results and analysis tive accuracy P achieved with three classifiers including
the novel K-RecSys–CF, SVM–CF and SVM–IACO–CF.
Predictive accuracy P, recall rate R and F-measure F of The predictive accuracy P results show that the proposed
the novel SVM–IACO based on filtering collaboration and SVM–IACO–CF classifier yields a much better predictive
other recommendation algorithms such as K-RecSys–CF, accuracy value of 20%, while the other available techniques
like K-RecSys–CF, SVM–CF, a lower predictive accuracy

13
Optimized machine learning based collaborative filtering (OMLCF) recommendation system…

45 in Fig. 5. Three classification techniques include K-Rec-

Sys–CF, SVM–CF and SVM–IACO–CF. The attained
40
results show that the newly introduced SVM–IACO–CF
35 classifier yields better F-measure results of 32%, while
other available techniques like K-RecSys–CF, SVM–CF
Recall rate R(%)

30
yields low results of 25% and 30% correspondingly. With
25 the value of recommended item number N being increased,
K-RecSys-CF there is also an increase in the F-measure for three diverse
20
SVM-CF classifiers.
15 SVM-IACO-CF

6 Conclusion and future work

10

5
The Internet’s fast-paced development and mobile Inter-
0
5 10 15 20 25 30 net have made the industry of e-commerce thriving on the
Recommended item number N widespread interest from every walk of life. There has been
tremendous pressure from e-commerce on the recommenda-
Fig. 4  Recall rate R of three models with different N values tion system to manage a massive number of customers and to
yield superior recommendation quality. Further, it is decided
by the recommended item’s count and the performance ren-
35 dered by the recommendation algorithm. Also, collaborative
filtering algorithms can fall into two groups, which include
30
item-based and user-based recommender systems. A chal-
25 lenge including poor recommendation predictive accuracy
F-measure F(%)

and effectiveness is frequently observed in classical collabo-

20
K-RecSys-CF rative filtering due to the massive elementary data volume.
15
SVM-CF Therefore, this research work has adopted SVM first for the
10 SVM-IACO-CF
items’ classification and then filtered out the user disliked
things using the information which has negative-feedback
5 to decrease the amount of recommended commodities.
0
Once the positive feedback items are filtered, scores and
5 10 15 20 25 30 comments are considered to get the elaborates cores using
Recommended item number N the weighted average, thereby rendering it more objective
and resulting in subtle predictive accuracy improvement.
Fig. 5  F-measure F of three models with different N values The issues involving lesser effectiveness in SVM classi-
fier has to be mitigated. The optimized values of the SVM
value of 15, 18% is achieved correspondingly. With the parameters are obtained through the IACO algorithm. At
increase in the value of recommended item number N, the last, SVM–IACO based collaborative filtering algorithm is
predictive accuracy P value is reduced for three diverse used. The recommendation of collaborative filtering helps
classifiers. only on commodities that have positive-feedback. Python-
The Recall rate R of three models having diverse N val- based Scrapy experiments were enforced for the data extrac-
ues were explained in Fig. 4 and further, it explains the tion of information that occurs in the Taobao dataset in the
Recall rate metrics of three classifiers including the newly online commodity. Tests carried out on Taobao data (an
introduced K-RecSys–CF, SVM–CF, and SVM–IACO–CF. Alibaba owned Chinese online shopping website) revealed
It is shown from the Fig. 2 it reveals that the newly intro- that the algorithm yields a superior recommendation pre-
duced SVM–IACO–CF classifier yields a greater recall rate dictive accuracy. The newly introduced SVM–IACO–CF
value of 39%, whereas the available techniques including system and the available techniques are measured in terms
K-RecSys–CF, SVM–CF yields low recall rate value of 35, of Predictive accuracy P, Recall rate R and F-measure F.
37% correspondingly. On increasing the value of the recom- For instance, the predictive accuracy P results show that the
mended item number, the recall rate R-value also sees an proposed SVM–IACO–CF classifier yields a much better
increase for three diverse classifiers. predictive accuracy value of 20%, while the other available
The results of the F-measure of the three classifica- techniques like K-RecSys–CF, SVM–CF, a lower predictive
tion techniques with diverse N values were demonstrated accuracy value of 15, 18% is achieved correspondingly. With

13
 J. Anitha, M. Kalaiarasu

the increase in the value of recommended item number N, Lin S, Wenzheng X (2015) E-commerce personalized recommendation
the predictive accuracy P-value is reduced for three diverse system based on web mining technology design and implementa-
tion. In: International conference on intelligent transportation, big
classifiers. More analysis can be conducted in the future on data and smart city, pp 347–350
the means of further improving the accuracy and compe- Lu PY, Wu XX, Teng DN (2015) Hybrid recommendation algorithm
tence of the algorithm which is recommended when we have for e-commerce website. In: International symposium on compu-
many. Also, we will try to optimize our algorithm in order tational intelligence and design (ISCID), pp 197–200
Prando AV, Contratres FG, de Souza SNA, de Souza LS (2017) Con-
to provide better accuracy with a high level of privacy and tent-based recommender system using social networks for cold-
implement in real time applications. start users. In: KDIR, pp 181–189
Ramzan B, Bajwa IS, Jamil N, Amin RU, Ramzan S, Mirza F, Sarwar
N (2019) An intelligent data analysis for recommendation systems
using machine learning. Sci Prog 2019(5941096):1–20
References Rehman A, Rathore MM, Paul A, Saeed F, Ahmad RW (2018) Vehicu-
lar traffic optimisation and even distribution using ant colony in
Chang D, Gui HY, Fan R, Fan ZZ, Tian J (2019) Application of smart city environment. IET Intel Transp Syst 12(7):594–601
improved collaborative filtering in the recommendation of Syarif I, Prugel-Bennett A, Wills G (2016) SVM parameter optimiza-
e-commerce commodities. Int J Comput Commun Control tion using grid search and genetic algorithm to improve classifica-
14(4):489–502 tion performance. Telkomnika 14(4):1–8
Che G, Liu L, Yu Z (2019) An improved ant colony optimization algo- Wang B, Ye F, Xu J (2018) A personalized recommendation algorithm
rithm based on particle swarm optimization algorithm for path based on the user’s implicit feedback in e-commerce. Future Inter-
planning of autonomous underwater vehicle. J Ambient Intell net 10(12):117–129
Human Comput. https​://doi.org/10.1007/s1265​2-019-01531​-8 Wasid M, Kant V (2015) A particle swarm approach to collaborative
Feng J, Fengs X, Zhang N, Peng J (2018) An improved collaborative filtering based recommender systems through fuzzy features. Pro-
filtering method based on similarity. PLoS ONE 13(9):1–18 cedia Comput Sci 54:440–448
Gohari FS, Haghighi H, Aliee FS (2017) A semantic-enhanced trust Wu X, Wei G, Song Y, Huang X (2018) Improved ACO-based path
based recommender system using ant colony optimization. Appl planning with rollback and death strategies. Syst Sci Control Eng
Intell 46(2):328–364 6(1):102–107
Hu ZH, Li X, Wei C, Zhou HL (2019) Examining collaborative filtering Yan L (2017) Personalized recommendation method for e-commerce
algorithms for clothing recommendation in e-commerce. Text Res platform based on data mining technology. In: International con-
J 89(14):2821–2835 ference on smart grid and electrical automation (ICSGEA), pp
Hwangbo H, Kim YS, Cha KJ (2018) Recommendation system devel- 514–517
opment for fashion retail e-commerce. Electron Commer Res Appl Zarzour H, Al-Sharif Z, Al-Ayyoub M, Jararweh Y (2018) A new col-
28:94–101 laborative filtering recommendation algorithm based on dimen-
Jaganathan S, Palaniswami S, Vignesh GM, Mithunraj R (2011) Appli- sionality reduction and clustering techniques. In: International
cations of multi objective optimization to reactive power planning conference on information and communication systems (ICICS),
problem using ant colony algorithm. Eur J Sci Res 51(2):241–253 pp 102–106
Jiang L, Cheng Y, Yang L, Li J, Yan H, Wang X (2019) A trust-based Zhang X, Chen X, He Z (2010) An ACO-based algorithm for param-
collaborative filtering algorithm for e-commerce recommenda- eter optimization of support vector machines. Expert Syst Appl
tion system. J Ambient Intell Humaniz Comput 10(8):3023–3034 37(9):6618–6628
Li X, Li D (2019) An improved collaborative filtering recommenda- Zhao X (2019) A study on e-commerce recommender system based on
tion algorithm and recommendation strategy. Mobile Inf Syst big data. In: IEEE 4th international conference on cloud comput-
2019(3560968):1–11 ing and big data analysis (ICCCBDA), pp 222–226
Li ZL, Hu CX, Wei XY, Zou TF, Zhang HR, Yang GC (2014) Enhanc-
ing collaborative filtering recommendation by utilizing improved Publisher’s Note Springer Nature remains neutral with regard to
ant colony optimization algorithm. Applied mechanics and mate- jurisdictional claims in published maps and institutional affiliations.
rials, vol 556. Trans Tech Publications Ltd., pp 3793–3799
Li ZX, Li C, Jue Z (2016) Multi-objective particle swarm optimi-
zation algorithm for recommender system. Adv Model Anal B
59(1):189–200

13