An Efficient Recommendation System in E-commerce using Passer learning

optimization based on Bi-LSTM

Hemn Barzan Abdalla1*, Awder Ahmed2, Bahtiyar Mehmed3, Mehdi Gheisari4, Maryam Cheraghy5
1,5 Department of Computer Science, Wenzhou-Kean University, Wenzhou, China
2 Department of Communication Engineering, Sulaimani Polytechnic University, Iraq
3 Department of Economics, Neusoft Institute Guangdong, Foshan, China
4 Department of Computer Science, Neusoft Institute Guangdong, Foshan, China

*Corresponding author: Hemn Barzan Abdalla; [email protected]

Abstract: Recommendation system services have become crucial for users to access person-
alized goods or services as the global e-commerce market expands. They can increase business
http://arxiv.org/abs/2308.00137

sales growth and lower the cost of user information exploration. Recent years have seen a sig-
nificant increase in researchers actively using user reviews to solve standard recommender sys-
tem research issues. Reviews may, however, contain information that does not help consumers
decide what to buy, such as advertising or fictitious or fake reviews. Using such reviews to offer
suggestion services may reduce the effectiveness of those recommendations. In this research,
the recommendation in e-commerce is developed using passer learning optimization based on
Bi-LSTM to solve that issue (PL optimized Bi-LSTM). Data is first obtained from the product
recommendation dataset and pre-processed to remove any values that are missing or incon-
sistent. Then, feature extraction is performed using TF-IDF features and features that support
graph embedding. Before submitting numerous features with the same dimensions to the Bi-
LSTM classifier for analysis, they are integrated using the feature concatenation approach. The
Collaborative Bi-LSTM method employs these features to determine if the model is a recom-
mended product. The PL optimization approach, which efficiently adjusts the classifier's pa-
rameters and produces an extract output that measures the f1-score, MSE, precision, and recall,
is the basis of this research's contributions. As compared to earlier methods, the proposed PL-
optimized Bi-LSTM achieved values of 88.58%, 1.24%, 92.69%, and 92.69% for dataset 1,
88.46%, 0.48%, 92.43%, and 93.47% for dataset 2, and 92.51%, 1.58%, 91.90%, and 90.76%
for dataset 3.

Keywords: Recommendation system; e-commerce; passer learning optimization; Bi-LSTM

classifier and TF-IDF.
 2

1. Introduction
Online shopping has expanded significantly and gained popularity in recent years due to wide-
spread internet use and the quick rise in mobile usage (Olagunju et al. , 2020). The practice of
purchasing, offering for sale, or exchanging goods, services, and information using computer
networks like the Internet is known as e-commerce (Shahjee, 2016). However, buyers must sort
through a deluge of information produced by modern e-commerce to decide which products to
buy. The Recommendation System can be used to address the issue of information overload.
Using e-commerce websites, the recommendation system is utilized to suggest products to its
customers (Gosh et al., 2020). The recommendation system benefits both the buyer and vendor
communities in an e-commerce environment. Recommendation systems are required to obtain
pertinent products and more individualized recommendations based on user purchasing behav-
ior and product interests from the vast number of products (Li et al. 2019; Karthik & Ganapathy,
2021). Recommendations have long been an important source of traffic for e-commerce com-
panies. In the past, the proposed scenarios on the e-commerce platform covered each stage of
the transaction, such as the information page, cart, order, and payment (Lange et al., 2015).
Product recommendations on e-commerce websites can enhance product sales and increase the
website's overall effective conversion rate for product sales. Users' records that they have
looked through, gathered, and purchased will help us better understand their needs over time,
categories and label users, recommend products they might be interested in, help users find the
products they need quickly, and sell a wider range of goods (Adhikary, 2014; Zhu et al.
2018;Zhou 2020) .
People can rate the things they are interested in at any time and from any location in online e-
commerce networks. The social connections between users and the things they have evaluated
can create a complicated network structure.
Traditional personalized suggestions used in heterogeneous e-commerce recommender sys-
tems, like collaborative filtering, content-based methods, and hybrid approaches, largely gather
semantic features based on interactions between users and items or behavioral history data. The
majority of studies disregard heterogeneous type data, which can be used to identify specific
user and item semantics and enhance recommender system quality (Zheng et al., 2021). Web
recommender systems (Moreno et al., 2016) inherited filtering methods to anticipate user needs
that user might have the propensity to express. Consider, for instance, the Netflix or Amazon
product display models, which suggest content to users based on their past behavior (Sharma et
al., 2022). In order to make it easier for consumers to choose products that suit their individual
 3

interests, recommender systems were developed in the late 1990s. Typically, recommender sys-
tems base their judgments on the opinions expressed by end users in the form of rating feedback.
Cross-domain recommender systems, which allow the transfer of data from a source domain
with dense ratings to a target domain with sparse ratings, have been created recently. Such
information transfer helps conventional recommender systems get over their problems with
cold starts and data scarcity (Khan et al., 2017; Cremonesi et al., 2011; Ahmed et al., 2021).
The Recommendation System utilizes various technologies to select the optimal outcome and
give consumers the required data. Recommendation systems can be broadly categorized into
three groups: collaborative, content, and hybrid recommendation systems (Alzogbi et al., 2019).
Content-based systems attempt to verify the item's suggested actions. This works by figuring
out how new users can behave depending on the knowledge they need, which is provided as
objects by the user (Verma et al., 2015).
Recommender Systems (RS) are frequently employed in digital platforms such as e-commerce,
video, and music platforms since they may assist in finding and recommending the most perti-
nent goods among vast alternative choices. RS are significant, but they encounter a number of
obstacles. Providing product recommendations, for instance, will be difficult in the e-commerce
space when a customer visits anonymously or without viewing any items, as little to no infor-
mation is available about the user (Esmeli et al., 2020). Massive product offerings have made
it difficult to maintain user satisfaction in contemporary online services like social media and
e-commerce. Recommender systems are one of the most important factors in reducing this in-
formation overload. They typically offer a select group of goods from among millions or even
billions of options (Xie et al., 2021). When users join e-commerce websites, such information
is already made available to the public. Making recommendations using auxiliary information
from users or product reviews has thus become a growing study area (Hong & Jung 2018; Zhou
& Zha, 2012). Traditionally, most current recommendation systems use item- or user-based
approaches to predict users' rating scores. These methods use calculated user similarity or item
similarity to forecast the scores of unrated things. However, these approaches always perform
poorly when there are a small number of scored objects. Furthermore, it is challenging to make
diverse suggestions (Zhang et al., 2020). The Hybrid Recommendation System fulfills its duties
through the use of content and collaborative filtering methodologies to evaluate whether a cer-
tain item is appropriate. Because it can eliminate any flaw that might have developed during
the implementation of a recommendation system, the hybrid system is regarded by many or-
ganizations as the most commonly used recommendation system (Chen et al., 2016).
 4

The primary goal of this research is to develop an e-commerce recommendation system em-
ploying passer learning optimization based on Bi-LSTM. The data is first gathered from the
recommendation dataset and pre-processed to remove any values that are missing or incon-
sistent. Then, feature extraction is performed using TF-IDF features and features that support
graph embedding. Before submitting numerous features with the same dimensions to the Bi-
LSTM classifier for analysis, the feature concatenation approach merges them. Finally, the
model’s collaborative bi-LSTM classifier predicts the products that have the required attributes.
The passer learning optimization method provides the extracted output, which successfully ad-
justs the classifier's parameters. The following are the research's main contributions:
Passer learning optimization (PL): The Passer learning optimization algorithm is designed by
integrating the automatic learning capacity of the teacher (Guo et al., 2022 and is merged with
the sparrow (Natarajan et al., 2022) to improve the convergence rate, speed, and performance.
PL-optimized Bi-LSTM: Bi-LSTM uses the PL optimization technique to modify the
classifier's parameters efficiently. The classifier's enhanced recommendations system in e-com-
merce: lower computing complexity, quicker convergence, and increased effectiveness, are all
outcomes of using the PL optimization method.
The manuscript is organized as follows; The proposed methodology for the e-commerce rec-
ommendation system is described in the second section. The suggested Passer learning optimi-
zation technique and related mathematical models are described in the third section. The exper-
imental findings of the suggested method are presented in the fourth section, and the conclusion
and suggestions for future work are provided in the fifth section.
Discussion now follows of existing models for e-commerce recommendation systems. A new
fuzzy logic-based product suggestion system was developed by (Karthik & Ganapathy, 2021)
that utilizes users' present interests to anticipate which products are the most pertinent to them
when they purchase online. This approach raised the diversity and product rating scores but
decreased the recommendation system's level of confidence. A deep learning and distributed
expression-based application solution for e-commerce product advertising recommendations
was created by (Zhou, 2020). This approach had great flexibility and minimized calculating
complexity, but it was exceedingly expensive and time-consuming. (Zheng et al., 2021) pro-
posed a heterogeneous product system that combined user, item, and interaction information.
This strategy increased quality and decreased rating prediction error but converged too slowly.
(Sharma et al., 2022) introduced a hybrid recommendation system that anticipates recommen-
dations. The suggested method combined collaborative filtering with content-based filtering.
Although this method's recommendations were more accurate, it had a cold start issue. In the
 5

cross-domain situation of user overlap in an e-commerce system, (Ahmed et al. ,2021) created
a domain matrix factorization that predicted the rating of a product for a frequent user with a
fixed cold start issue. Although this method's convergence is excessively sluggish and does not
predict all testing data, it does minimize the time complexity. (Esmeli et al., 2020) presented
the session similarity-based strategy to address the issue of cold-start sessions in the e-com-
merce area, where no interactive objects in the sessions can aid in discovering users' prefer-
ences. This approach lessened the effects of the cold-start movie issue. Self-complementary
collaborative filtering, a framework proposed by (Xie et al. , 2021), can make recommendations
in real time using both global and local information. This technique lowered the size of the
model and helped with overfitting, but it is time-consuming and cannot be used in large-scale
applications. (Zhang et al., 2020) presented a novel strategy that is effectively a hybrid proba-
bilistic matrix model. Although this method had a high computational complexity and a high
computing cost, it minimized errors. (Liu et al., 2021) proposed a Time Preference Gate net-
work model to predict the user’s purchase intentions; this model increases the performance of
intense prediction but it consumes more time. (Pan & Zhang, 2021) developed a context-aware-
ness mobile tourism E-commerce personalized recommendation model which predicts the pref-
erence and interest of the user with specific context, but although it achieves a high level of
user satisfaction, the efficiency of the algorithm is very low. (Jiang et al., 2019) presented a
slope one algorithm based on the fusion of trusted data and user similarity to solve low accuracy
problems in the traditional slope algorithm. Whilst it achieves high accuracy, it fails to solve
the cold start problem because no user preference information is available.
The following is a list of the difficulties encountered during the current research:
 Finding the best product among the many recognized possible products remains difficult.
 The proposed recommendation system's capacity to gather and consider users' evolving in-
terests is one of its biggest challenges.
 The commodities recommendation system is currently dealing with a number of new diffi-
culties, chief among them the diversification of the outcomes of the recommendations, the
personalization, and the timeliness of intelligent terminal recommendations.
 Collaborative filtering faces difficulties with scalability, data sparsity, and cold start.
 Information filtering is a significant difficulty because it provides for delivering desired
content that is highly relevant to user preferences.
 An issue incurred by having fewer total user ratings for a larger number of things is data
sparsity. One of the major problems facing recommender systems is the cold start issue.
 6

2. Materials and Methods

Next, the materials and methods used for the recommendation system in e-commerce are de-
scribed in detail.

2.1 Proposed Methodology for Recommendation System in E-commerce

The primary goal of the research is to create an e-commerce recommendation system employing
a Passer learning optimization-based Bi-LSTM classifier. Initially, the recommendation dataset
is inputted and then the data is pre-processed to enhance its quality. The pre-processed data is
forwarded to content and collaborative-based feature extraction, where the TF-IDF and the
graph embedding-enabled features are extracted. The graph embedding enables extraction of
the data features in the form of a discrete graph, which is represented in vector form, and the
TF-IDF features measure the relationship between Word and the document. Finally, from the
extracted features, the collaborative Bi-LSTM classifiers recommend a valid product to the
user. Passer learning optimization, which was created using two common optimization tech-
niques, namely teaching learning-based optimization (TLBO) (Natarajan et al., 2022) and Spar-
row search algorithm (SSA) (Guo et al., 2022), is used to improve the performance of the Bi-
LSTM classifier. Figure 1 depicts a schematic illustration of the suggested recommendation
system in an e-commerce model.

Figure 1: Schematic representation of the proposed recommendation system in the e-commerce

model
 7

2.2 Input
The input given for the recommendation model is mathematically formulated as the following
equation:

a
R   Ri (1)
i 1

where, the recommendation dataset is denoted as R , and Ri denotes images present in the da-

taset, which ranges from 1 to a .

2.3.Data pre-processing
Pre-processing is used to remove unnecessary information from the data, such as fixing incon-
sistencies, adding up any missing numbers, standardizing the data, and improving the quality
of the data. Before feeding text data to a classifier model, data pre-processing is used to clean
up the text data, which contains various types of noise, including emotions, punctuation, and
text in a different case. The pre-processed data is mathematically represented as:

a
R   Ri* (2)
i 1

where, the pre-processed image is denoted as Ri* .

2.4 Content & Collaborative Filtering-based Feature Extraction

An important phase of feature selection is eliminating unnecessary features, which avoids over-
fitting problems by lowering the size of the features. In order to accurately categorize the po-
larity, the algorithm is trained for a certain feature that is utilized to represent the data. The class
attribute is thus represented in the smaller feature space by feature selection, which chooses the
fewest significant characteristics. Techniques for feature selection can considerably increase
classification accuracy and give users a better knowledge of key class traits, which helps to
interpret the data. The two steps used in this research to extract the features: TF-IDF features
and Graph embedding enabled features, are each described and discussed in depth below.

2.4.1 TF-IDF features

The term ‘frequency-inverse document frequency’ (TF-IDF) is one of the most popular
weighting metrics for detecting the link between words and documents. It is commonly used in
the process of collecting word features. We can quantify the importance of every phrase in a
document by assigning it a numerical value with the aid of TF-IDF. The inverse document rate,
or IDF, quantifies the frequency with which words appear across the whole corpus. If a term
appears only sometimes in the corpus, it can still be considered a feature because it has great
 8

differentiation potential (Hong & Jung 2018). The equation for the TF-mathematical IDF is
calculated as follows:

Oi , j
Vti , j  (3)
 a
Oa , j

|Q|
idt i  log (4)
| j : wi  v j |

The denominator of the formula (3) is the total number of times all the words appear in the text,
and the value in the numerator of the formula is the number of times each word appears in the
document. The formula (4) | Q | shows the overall amount of documents in the corpus or the

entire number of documents comprising the phrases. If the feature word's TF  IDFt1 value

equals the word's value when Vti , j is multiplied by idti , the weight of the feature word can also

be determined using formula (5).

TF  IDFt1  vti , j  idti (5)

2.4.2. Graph embedding enabled features

Graph embedding refers to the processes that convert property graphs into a vector or a collec-
tion of vectors. Graph topology, vertex-to-vertex relationships, and other pertinent details about
graphs, subgraphs, and vertices should all be captured during embedding.
Because every node's attributes are compressed in a vector with a smaller dimension via graph
embedding techniques, node similarity in the original complicated irregular spaces may be read-
ily measured in the embedded vector spaces using common metrics.
Benefits of graphical representation include time savings and easier understanding of data. This
graph embedding feature for datasets 1, 2, and 3 is shown in Figure 2.

(a) (b)
 9

(c)

Figure 2: Graph embedding-based features for (a) dataset 1, (b) Dataset 2, and (c) Dataset 3

2.5 Feature concatenation

After extracting numerous features with the same dimensions to the Bi-LSTM classifier for
analysis, they are merged using the feature concatenation approach. The various original infor-
mation is significantly conserved throughout the concatenation method due to the integrated
perspective representation of all the concatenated features. The Passer learning optimization-
based Bi-LSTM classifier receives the output of feature concatenation and improves the per-
formance during the recommendation system. The optimization method discovers the most ef-
fective solution, while the overall number of parameters in the features is minimized. As a
result, the model's performance is assessed and trained, using both the test data and the trained
data as modeling input. A training dataset is used to train the model with the dimension of
1 2000 , and the accuracy of the model determines the test data.

2.6 Collaborative Bi-LSTM based recommendation model

The event recognition module receives its input from the feature concatenation, allowing for
additional data training in event detection and improving prediction accuracy over existing
methods. The training process is supported by optimization in addition to the feature vector,
and optimization is essential in reducing the danger of computational complexity and over-
fitting issues. Architecturally, one LSTM takes the input forward while the second LSTM takes
it backward by essentially increasing the network's information source. Finally, Bi-LSTM re-
verses the information flow by adding an extra LSTM layer. In other words, the input sequence
can flow backward by using a second LSTM layer. The two LSTM layers' outputs are then
combined using a range of operations, including averaging, adding, multiplying, and concate-
nating.
 10

Figure 3 displays the Bi-LSTM classifier recommendation system. This architecture is com-
prised of input layers, forwarding layers, backward layers, totally connected layers, and output
layers. A recurrent neural network (RNN), also known as the LSTM classifier, was created to
handle sequential data by assigning weights among the data. The Bi-LSTM classifier effectively
manages the error gradient by making gates available and giving long-term dependencies, sig-
nificantly reducing the vanishing gradient problem. The equation demonstrates the way the Bi-
LSTM classifier is conceptualized mathematically.

BX  Awc .xY  OY .cY 1  bc  (6)

The features expressed as Ri*, f are carried by the current word embedding represented by BX .

The weights for the Bi-LSTM classifier are given by wc and OY , whereas A X  stands for the

nonlinear function, bc determines the bias, and cY 1 signifies the hidden state. The classifier's
weights and bias are optimized using the Passer learning optimization, which also successfully
adjusts the hyper parameters.
EY   wE . yY  OE . cY 1  bE  (7)

J Y   wJ . yY  OJ .cY 1  bJ  (8)

PY   wP . yY  OP .cY 1  cP  (9)

mY  EY .mY 1  J Y . tanhwm .xY  Om .cY 1  cm  (10)

cY  PY . tanh mX  (11)

Here, EY , J Y and PY stand for the input gate, forget gate, and memory cell. mY stands for the
memory cell, PY for the Hadamard product, and  for the sigmoid function. While the forget
gate assists in forgetting previous knowledge, the memory cell in the input gate keeps the cur-
rently important information. The output gate determines which information is presented in the
internal memory cell and allows for many retrievals of crucial information.
 11

Figure 3: Bi-LSTM architecture

2.7. Proposed Passer Learning Optimization

In order to maximize the efficiency of the Bi-LSTM classifier, an effective passer learning op-
timization is enabled for the tuning of the parameters. The passer learning optimization algo-
rithm is developed by the combination of the characteristics of sparrow (Natarajan et al., 2022)
and teacher learning (Guo et al., 2022) optimizations, which combines the automatic learning
features of the teacher with sparrow to increase convergence rate, speed, and performance. A
detailed explanation of the algorithm with mathematical notations is described in the below
sections.

2.7.1 Motivation
Sparrows are quite bright birds with good memory capabilities, and based on their intelligence,
the sparrows are organized into three categories: producer, scrounger, and investigator. The
producers frequently have abundant energy reserves and provide access to foraging locations
or directions to all scavengers. Locating areas with an abundance of food supplies is the respon-
sibility of the producers. The degree of an individual's energy reserves is assessed based on
their fitness ratings. Additionally, as soon as they notice the predator, the sparrows begin chirp-
ing as an alarm signal. If the alarm value surpasses the safety level, the producers are required
to direct all scavengers to the safe area. Every sparrow has the capacity to produce as long as it
looks for better food sources, but the population as a whole still has the same proportion of
scavengers as producers. The producers would be the more active sparrows. In order to gain
energy, many starving scavengers are more likely to fly to various regions in quest of food.
Scroungers follow the producer who can provide the best food as they search for food. Some
scavengers may constantly be monitoring the producers and contending for them to up their
 12

predation rate while they wait for food. While the sparrows in the core of the group flit aimlessly
to stay close to one another, the sparrows at the group's boundaries migrate quickly towards the
safe area to take up a better position when in danger. Simultaneously, the teaching-learning-
based optimization algorithm (TLBO) is a unique population-based meta-heuristic intelligence
algorithm, which is motivated by the teacher's teachings and the students' performance in a
class. The Teacher and Learner phases are the two most important components of the algorithm.
During the teacher phase, students learn from their teachers. The instructor, who puts much
effort into raising students to his or her level, is thought of as the class's most knowledgeable
individual. During the learning phase, students advance their knowledge through sharing
knowledge. A learner can increase his or her knowledge through casual interactions with other
students, such as group discussions, formal presentations, and communications. Additionally,
a learner gets knowledge from a more experienced and knowledgeable person.

2.7.2 Mathematical model of the proposed Passer learning optimization

The producer, scrounger, and investigator are the three categories used to represent the mathe-
matical model of the proposed Passer learning optimization, and those behaviors are mathemat-
ically formulated in the below sections.
Position update of producer
Producers in Sparrow are responsible for finding food and controlling the movement of the
entire population, though producers with higher fitness ratings have priority when it comes to
finding food. As a result, the producers have access to a greater range of food sources than the
scrounger. When the sparrows see the predator, they immediately begin chirping as an alert. If
the alarm value beats the safety level, the producers are required to direct all scavengers to the
safe area. The location update producer's mathematical formula is as follows:
 t  h 
 x  ex 
  M ijt  xijt  R2  AP
xijt 1    1  M
ij
 (12)

 xt  r * fl 
 ij i   
M ijt  xijt R2  AP

The current iteration is denoted as h , the random number is denoted as 1 which ranges from

0 to 1, the flight length is denoted as fl , the uniform random number is denoted as ri which
ranges from 0 to 1, the maximum number of iterations is denoted as M , the alarm value is
denoted as R2 , awareness probability is denoted as AP , and where xijt represents the current

position of the j th dimension of the i th sparrow at the iteration t . when R2  AP , indicating

 13

that there are no nearby predators. All sparrows must immediately fly to other safe regions if
R2  AP , which indicates that some sparrows have identified the predator.

Position update of scrounger

Some scavengers keep a closer eye on the producers and engage in feeding competition with
them to increase the pace of their own predation. As soon as they find out that the producer has
located delicious food, they promptly leave their current location to compete for food. If they
are successful, they can get the food right away; if not, the process continues. The scrounger's
position update formula is described as follows:
  M ijt 1  xijt   x tp  xijt 
Q  ex   rand  
t 1   i 2 
   
 f x t 1
ij


xij   (13)
 t  xijt  M ijt 
    Q  vijt


x ij k
    
f x f x 
t
ij
t
ij
1

where Q is a random number that follows the normal distribution, the producer's best solution

is denoted as x tp , the i th sparrow's velocity in the j th dimension at iteration t is denoted as

vijt , the sparrow's direction of motion and the step size control coefficient are both denoted as

k , and the fitness functions of the current iteration, the next iteration, and the previous iteration
 
are denoted as f xijt , f xijt 1  and f xijt 1  .

Position update of investigators

The investigators in Sparrow are chosen at random from the population. When predators enter,
investigators send out signals to cause sparrows to run to a safe location, but because the spar-
row's automatic predicting capacity is not interpreted, the performance slows down and results
in a poor convergence rate. The automatic learning capability of the teacher learning is com-
bined with the sparrow in order to enhance convergence rate and performance so that it can
independently identify the predator. The global convergence is increased as a result, and the
time complexity is minimized. The combined algorithm's mathematical formula is:
xijt 1 | x |  A  L  M ijt (14)

1 (15)
xijt 1  xijt  xijt 1  | A  L  M ijt |
2
where A stands for a 1 d matrix, each of whose elements is given a random value of 1 or 1,

and A  AT  AAT  . L displays a matrix of 1 d whose elements are all 1. The Pseudo code
1

for the proposed Passer learning optimization is provided in Table 1.

 14

Table 1: Pseudo code for the proposed Passer learning optimization

S.No Pseudo code for the proposed Passer learning optimization

1. Initialization

2. M : maximum number of iterations

3. R2 : alert value

4. Q : random value

5. Initialize population

6. t 1

7. While t  M 
8. Update the position of the producer

 t  h 
9.
 x  ex 
  M ijt  xijt  R2  AP
xijt 1    1  M
ij


 xt  r * fl 
 ij i   
M ijt  xijt R2  AP

10. R2  rand 1

11. for i  1

12. Update the position of the scrounger

13.   M ijt 1  xijt   x tp  xijt 


Q  ex   rand  
t 1   i 2 

 f x t 1
 ij   

xij  
 t  xijt  M ijt 
    Q  vijt

x
 ij k
    
f x f x 
t
ij
t
ij
1


14. End for

15. for i  N  1

16. Update the position of the investigators

1
17. xijt 1  xijt  xijt 1  | A  L  M ijt |
2
18. End for

19. End while

 15

3. Result
The results derived from employing the PL optimized Bi-LSTM for the e-commerce recom-
mendation system are discussed in detail in the section below.

3.1 Experimental Setup

In order to compare the performance of the proposed and existing approaches for e-commerce
recommendations, the PL-optimised Bi-LSTM classifier is implemented in a PYTHON pro-
gram on Windows 10 with 8 GB RAM.

3.2 Dataset description(Product recommendation dataset)

Product reviews and metadata from Amazon are included in the recommendation collection.
These reviews total 143.7 million and cover the period from May 1996 to July 2014. Reviews,
product metadata, and links are all included. Three datasets, including the baby, digital music,
and patio lawn garden datasets, are used in this research.

3.3 Parameter metrics

3.3.1 Precision
Precision is defined as the ratio of correctly predicted positive results to all positively predicted
by the PL-optimized Bi-LSTM.
C pos (16)
C pre 
C pos  C pos

3.3.2 Recall
Recall is calculated by dividing the sum of true positives and false negatives by the total number
of true positives.
C pos (17)
Crecall 
C pos  Cneg

3.3.3 F1 score
The F1 score is the harmonic mean between recall and precision, and it is represented as fol-
lows:
2C pos (18)
C f 1 score 
2C pos  Cneg  C pos

3.3.4 Mean Squared Error (MSE)

When comparing real and estimated data, the difference is known as MSE, or mean square
error.
 16

3.4 Performance evaluation

For testing the classifier's performance during different epochs, the PL-optimized Bi-LSTM
classifier is put through a performance evaluation, detailed in the section below.

3.4.1 Performance evaluation-based training percentage for Dataset 1

Figure 4 displays the performance analysis of the suggested PL-optimized Bi-LSTM classifier
for dataset 1 for the training percentages of 40, 50, 60, 70, 80, and 90 for the varied epochs of
10, 15, 20, 25, and 30. The suggested PL-optimized Bi-LSTM's f1-score results are shown in
Figure 3a and were 66.43%, 73.83%, 75.94%, 78.18%, and 88.58% for a 90% training percent-
age. In a manner similar to this, the MSE of the proposed PL-optimized Bi-LSTM, as shown in
Figure 4b, attained 1.66%, 1.62%, 1.54%, 1.36%, and 1.24% for 90% training. For the 90%
training percentage given in Figure 4c, the precision of the suggested PL-optimized Bi-LSTM
was 66.74%, 75.07%, 77.89%, 79.43%, and 92.69%. Finally, the proposed PL-optimized Bi-
LSTM demonstrated in Figure 4d achieved 56.31%, 79.74%, 80.93%, 90.93%, and 92.69% for
a training percentage of 90% while evaluating recall.

(a) (b)

(c) (d)

Figure 4: Performance evaluation training percentage for dataset 1 a) f1score b) MSE c)

precision d) recall
 17

3.4.2 Performance evaluation-based training percentage for Dataset 2

Figure 5 displays the performance analysis of the suggested PL-optimized Bi-LSTM classifier
for dataset 2 for the training percentages of 40, 50, 60, 70, 80, and 90 for the varied epochs of
10, 15, 20, 25, and 30. The proposed PL-optimized Bi-LSTM's f1-score results are shown in
Figure 5a and were 80.72%, 83.21%, 83.66%, 86.55%, and 88.46%, with a 90% training rate.
Similar to this, as shown in Figure 5b, the MSE of the proposed PL-optimized Bi-LSTM
reached 0.61%, 0.60%, 0.53%, 0.50%, and 0.48% for 90% training. For the 90% training per-
centage given in Figure 5c, the precision of the suggested PL-optimized Bi-LSTM was 74.86%,
80.61%, 85.12%, 86.97%, and 92.43%. Finally, the proposed PL-optimized Bi-LSTM demon-
strated in Figure 5d achieved 75.83%, 90.26%, 90.47%, 91.74%, and 93.47% for training per-
centage of 90% while assessing recall.

(a) (b)

(c) (d)

Figure 5: Performance evaluation training percentage for dataset 2 (a) f1score (b) MSE
(c) precision (d) recall

3.4.3 Performance evaluation-based training percentage for Dataset 3

Figure 6 displays the performance analysis of the recommended PL-optimized Bi-LSTM clas-
sifier for dataset 3 for the training percentages of 40, 50, 60, 70, 80, and 90 for the varied epochs
 18

of 10, 15, 20, 25, and 30. The proposed PL-optimized Bi-LSTM's f1-score results are shown in
Figure 6a and were 89.15%, 89.23%, 90.35%, 91.91%, and 92.51% for a training percentage
of 90%. Similar to this, as shown in Figure 6b, the MSE of the proposed PL-optimized Bi-
LSTM reached 2.32%, 2.16%, 2.08%, 1.77%, and 1.58% for 90% training. For the 90% training
percentage given in Figure 6c, the precision of the suggested PL-optimized Bi-LSTM was
82.77%, 85.69%, 86.67%, 89.95%, and 91.90%. Finally, the proposed PL-optimized Bi-LSTM
demonstrated in Figure 6d achieved 60.54%, 74.95%, 75.52%, 78.96%, and 90.76% for a train-
ing percentage of 90%.

(a) (b)

(c) (d)

Figure 6: Performance evaluation training percentage for dataset 3 (a) f1score (b) MSE
(c) precision (d) recall

3.5 Comparative methods

The proposed PL-optimized Bi-LSTM classifier is compared with the Hybrid Probabilistic Ma-
trix Factorization Model (HPMFM) [KB-1](Wang et al. 2022), Self-Complementary Collabo-
rative Filtering (SCCF) [KB-2](Hoshika et al. 2022), A3NCF Model [KB-3](Liu et al. 2022),
Bi-LSTM Model [KB-4](Roy et al. 2022), Bi-LSTM with Harris Hawk Optimization [KB-
 19

5](Kavitha et al. 2022), and Bi-LSTM with Particle Swarm Optimization (PSO)[KB-6](Zhang
et al. 2022). The results of the comparative methods are discussed in the below section.

3.5.1 Comparative Analysis-based Training Percentage for Dataset 1

The proposed PL-optimized Bi-LSTM classifier's performance is compared in Figure 7 for da-
taset 1 using the f1score, MSE, precision, and recall metrics. The improvement value obtained
by assessing the f1 score with a training percentage of 90% when comparing the suggested PL-
optimized Bi-LSTM approach to the Bi-LSTM with the PSO method shown in Figure 7a is
4.24. When compared to the Bi-LSTM with PSO method shown in Figure 7b, the recommended
PL-optimized Bi-LSTM approach achieves an improvement value of training percentage of
90% of 1.81 in terms of MSE. Comparing the suggested PL-optimized Bi-LSTM to the Bi-
LSTM with the PSO approach shown in Figure 7c, an enhanced value of training percentage of
90% is achieved, which equals 11.06. When compared to the Bi-LSTM with the PSO approach
shown in Figure 7d, the recommended PL-optimized Bi-LSTM method achieves an improve-
ment value of measuring recall with a training percentage of 90% of 9.13.

(a) (b)

(c) (d)

Figure 7: Comparative evaluation training percentage for dataset 1 (a) f1score (b) MSE
(c) precision (d) recall
 20

3.5.2 Comparative analysis based on training percentage for Dataset 2

Figure 8 compares the performance of the proposed PL-optimized Bi-LSTM classifier for da-
taset 2 using the f1score, MSE, accuracy, and recall metrics with respect to the proportion of
training data. The improvement value obtained by calculating the f1 score with a training per-
centage of 90% when comparing the suggested PL-optimized Bi-LSTM approach to the Bi-
LSTM with the PSO method shown in Figure 8a is 1.66. When compared to the Bi-LSTM with
PSO method shown in Figure 8b, the recommended PL-optimized Bi-LSTM approach achieves
an improvement value of training percentage of 90% of 11.92 in terms of MSE. Comparing the
suggested PL-optimized Bi-LSTM to the Bi-LSTM with the PSO approach shown in Figure 8c,
an enhanced value of training percentage of 90% is achieved, which is 3.06. When compared
to the Bi-LSTM with PSO method shown in Figure 8d, the recommended PL-optimized Bi-
LSTM approach achieves an improvement value of measuring recall with a training percentage
of 90% of 5.76.

(a) (b)

(c) (d)

Figure 8: Comparative evaluation training percentage for dataset 2 (a) f1score (b) MSE
(c) precision (d) recall

3.5.3 Comparative analysis based on training percentage for Dataset 3

 21

Figure 9 compares the performance of the proposed PL-optimized Bi-LSTM classifier for da-
taset 3 using the f1score, MSE, accuracy, and recall metrics with respect to the proportion of
training data. The improvement value obtained by assessing the f1 score with a training per-
centage of 90% when comparing the suggested PL-optimized Bi-LSTM approach to the Bi-
LSTM with the PSO method shown in Figure 9a is 5.91. When compared to the Bi-LSTM with
PSO method shown in Figure 9b, the recommended PL-optimized Bi-LSTM approach obtains
an improvement value of training percentage of 90%, which is 4.68 in terms of MSE. Compar-
ing the suggested PL-optimized Bi-LSTM to the Bi-LSTM with the PSO approach shown in
Figure 9c, a better value of training percentage of 90% is obtained, which is 5.65. When com-
pared to the Bi-LSTM with PSO method shown in Figure 9d, the recommended PL-optimized
Bi-LSTM approach achieves an improvement value of measuring recall with a training percent-
age of 90% of 1.47.

(a) (b)

(c) (d)

Figure 9: Comparative evaluation training percentage for dataset 3 (a) f1score (b) MSE
(c) precision (d) recall

4 Comparative discussion
 22

The best results obtained are interpreted in Table 2 and compared with the proposed PL-opti-
mized Bi-LSTM approach. Due to the Passer learning optimized Bi-LSTM classifier being
enabled by strengthening the sparrow and instructor learning characteristics, the approach
achieved good metrics values when compared to other methods. The improvement of these
characteristics aids in effective tuning of the Bi-LSTM classifier to produce the required re-
sults.

Table 2: Comparative discussion of the proposed PL-optimized Bi-LSTM method

Method
KB - KB - KB - KB - KB - KB - Proposed
Dataset s/
1 2 3 4 5 6 Method
Metrics

F1-
66.91 68.25 68.44 86.25 86.43 86.87 90.72
score
Dataset
MSE 2.52 2.47 2.42 1.34 1.25 1.24 1.22
1
Precisio
57.66 57.89 58.81 79.90 82.25 83.19 93.52
n

Recall 74.70 75.64 83.37 83.83 83.89 84.10 92.54

F1-
56.71 59.07 60.25 80.95 83.86 88.93 90.43
score
Dataset
MSE 1.80 1.74 1.71 0.77 0.63 0.59 0.53
2
Precisio
52.63 56.61 57.74 85.83 87.55 92.10 95.00
n

Recall 61.44 62.67 66.16 75.57 86.33 88.78 94.20

F1-
61.71 62.21 67.27 79.55 83.83 88.67 94.24
Dataset score

3 MSE 2.41 2.36 2.05 1.73 1.62 1.47 1.41

Precisio
53.59 54.66 60.00 74.69 78.81 90.69 96.12
n

Recall 56.18 57.31 59.06 73.10 80.60 92.00 93.38

 23

5. Conclusion and future work

This research has developed the e-commerce recommendation system using Passer learning
optimization based on Bi-LSTM. The information is gathered from the product suggestion da-
taset, and pre-processing is done to obtain and get rid of any inconsistent or missing data values.
Then, TF-IDF features and features that support graph embedding are used to extract features.
Before supplying them to the Bi-LSTM classifier, the feature concatenation approach is applied
for multiple features with the same dimensions. Finally, the Collaborative Bi-LSTM classifier
is fed with these attributes in order to determine whether the model is a recommended product.
The PL optimization technique efficiently adjusts the classifier's parameters and produces an
extract output that measures the f1-score, MSE, precision, and recall. When compared to earlier
techniques, the suggested PL-optimized Bi-LSTM achieved values of 88.58%, 1.24%, 92.69%,
and 92.69% for dataset 1, 88.46%, 0.48%, 92.43%, and 93.47% for dataset 2, and 92.51%,
1.58%, 91.90%, and 90.76% for dataset 3. Future studies will focus on the hybrid models of
recommendation systems.

References
Adhikary, A. (2014). Advertising: A fusion process between consumer and product. Economics
and Finance, Procedia 11: 230-238. DOI 10.1016/S2212-5671(14)00191-9
Ahmed, A., Saleem, K., Khalid, O., & Rashid, U. (2021). On deep neural network for trust
aware cross domain recommendations in E-commerce. Expert Systems with Applications,
174: 114757. doi.org/10.1016/j.eswa.2021.114757
Alzogbi, A., Koleva, P., & Lausen, G. (2019). Towards distributed multi-model learning on
apache spark for model-based recommender. In 2019 IEEE 35th International Conference
on Data Engineering Workshops (ICDEW), 193-200. IEEE.
https://doi.org/10.1109/ICDEW.2019.00-12

Chen, Y. C., Yang, C. C., Liau, Y. J., Chang, C. H., Chen, P. L., Yang, P. C., & Ku, T. (2016).
User behavior analysis and commodity recommendation for point-earning apps. In 2016
Conference on Technologies and Applications of Artificial Intelligence (TAAI), 170-177.
IEEE. https://doi.org/10.1109/TAAI.2016.7880109

Cremonesi, P., Tripodi, A., & Turrin, R. (2011). Cross-domain recommender systems. In
2011 IEEE 11th International Conference on Data Mining Workshops, 496-503.
DOI: 10.1109/ICDMW.2011.57
 24

Esmeli, R., Bader-El-Den, M., & Abdullahi, H. (2020). Session Similarity based Approach for
Alleviating Cold-start Session Problem in e-Commerce for Top-N Recommendations. In
KDIR,. 179-186. DOI:10.5220/0010107001790186
Gosh, S., Nahar, N., Wahab, M. A., Biswas, M., Hossain, M. S., & Andersson, K. (2020). Rec-
ommendation system for e-commerce using alternating least squares (ALS) on apache
spark. In International Conference on Intelligent Computing & Optimization, 880-893.
Cham: Springer International Publishing. DOI.org/10.1007/978-3-030-68154-8_75
Guo, J., Yang, L., Rincón, D., Sallent, S., Chen, Q., & Liu, X. (2022). Static Placement and
Dynamic Assignment of SDN Controllers in LEO Satellite Networks. IEEE Transactions
on Network and Service Management. DOI 10.1109/TNSM.2022.3184989
Hong, M., & Jung, J. J. (2018). Multi-sided recommendation based on social tensor factoriza-
tion. Information Sciences, 447, 140-156. DOI.org/10.1016/j.ins.2018.03.019
Hoshika, S., Shukla, M. S., Benner, S. A., & Georgiadis, M. M. (2022). Visualizing “Alterna-
tive Isoinformational Engineered” DNA in A-and B-Forms at High Resolution. Journal of
the American Chemical Society, 144(34): 15603-15611. DOI.org/10.1021/jacs.2c05255
Jiang, L., Cheng, Y., Yang, L., Li, J., Yan, H., & Wang, X. (2019). A trust-based collaborative
filtering algorithm for E-commerce recommendation system. Journal of ambient intelli-
gence and humanized computing, 10: 3023-3034. DOI.org/10.1007/s12652-018-0928-7
Karthik, R. V., & Ganapathy, S. (2021). A fuzzy recommendation system for predicting the
customers interests using sentiment analysis and ontology in e-commerce. Applied Soft
Computing, 108: 107396. DOI.org/10.1016/j.asoc.2021.107396
Kavitha, R. J., Thiagarajan, C., Priya, P. I., Anand, A. V., Al-Ammar, E. A., Santhamoorthy,
M., & Chandramohan, P. (2022). Improved Harris Hawks Optimization with Hybrid Deep
Learning Based Heating and Cooling Load Prediction on residential buildings. Chemo-
sphere, 309: 136525. DOI.org/10.1016/j.chemosphere.2022.136525
Khan, M. M., Ibrahim, R., & Ghani, I. (2017). Cross domain recommender systems: a system-
atic literature review. ACM Computing Surveys (CSUR), 50(3): 1-34.
DOI.org/10.1145/3073565
Lange, T., Hoefges, M., & Ribisl, K. M. (2015). Regulating tobacco product advertising and
promotions in the retail environment: a roadmap for states and localities. The Journal of
Law, Medicine & Ethics, 43(4): 878-896. DOI.org/10.1111/jlme.12326
 25

Li, M., Wu, H., & Zhang, H. (2019). Matrix factorization for personalized recommendation
with implicit feedback and temporal information in social ecommerce networks. IEEE
Access, 7: 141268-141276. DOI:10.1109/ACCESS.2019.2943959
Liu, F., Cheng, Z., Chen, H., Wei, Y., Nie, L., & Kankanhalli, M. (2022). Privacy-preserving
synthetic data generation for recommendation systems. In Proceedings of the 45th Inter-
national ACM SIGIR Conference on Research and Development in Information Retrieval,
1379-1389. DOI.org/10.1145/3477495.3532044
Liu, Y., Tian, Y., Xu, Y., Zhao, S., Huang, Y., Fan, Y., & Guo, P. (2021). TPGN: a time-
preference gate network for e-commerce purchase intention recognition. Knowledge-
Based Systems, 220: 106920. DOI.org/10.1016/j.knosys.2021.106920
Moreno, M. N., Segrera, S., López, V. F., Muñoz, M. D., & Sánchez, Á. L. (2016). Web mining
based framework for solving usual problems in recommender systems. A case study for
movies ‫׳‬recommendation. Neurocomputing, 176: 72-80. DOI.org/10.1016/j.neu-

com.2014.10.097
Natarajan, E., Kaviarasan, V., Ang, K. M., Lim, W. H., Elango, S., & Tiang, S. S. (2022).
Production wastage avoidance using modified multi-objective teaching learning based
optimization embedded with refined learning scheme. IEEE Access, 10, 19186-19214.
DOI: 10.1109/ACCESS.2022.3151088
Olagunju, T., Oyebode, O., & Orji, R. (2020). Exploring key issues affecting african mobile
ecommerce applications using sentiment and thematic analysis. IEEE Access, 8: 114475-
114486. DOI: 10.1109/ACCESS.2020.3000093
Pan, H., & Zhang, Z. (2021). Research on context-awareness mobile tourism e-commerce per-
sonalized recommendation model. Journal of Signal Processing Systems, 93: 147-154.
DOI.org/10.1007/s11265-019-01504-2
Roy, D. K., Sarkar, T. K., Kamar, S. S. A., Goswami, T., Muktadir, M. A., Al-Ghobari, H. M.,
& Mattar, M. A. (2022). Daily Prediction and Multi-Step Forward Forecasting of Refer-
ence Evapotranspiration Using LSTM and Bi-LSTM Models. Agronomy, 12(3): 594.
DOI.org/10.3390/agronomy12030594
Shahjee, R. (2016). The impact of electronic commerce on business organization. Scholarly
Research Journal for interdisciplinary studies, 4(27): 3130-3140.
DOI.org/10.1109/HICSS.1997.663362
 26

Sharma, S., Rana, V., & Malhotra, M. (2022). Automatic recommendation system based on
hybrid filtering algorithm. Education and Information Technologies, 1-16.
DOI.org/10.1007/s10639-021-10643-8
Verma, J. P., Patel, B., & Patel, A. (2015). Big data analysis: recommendation system with
Hadoop framework. In 2015 IEEE International Conference on Computational Intelli-
gence & Communication Technology. 92-97. IEEE. DOI: 10.1109/CICT.2015.86
Wang, Z. C., Ran, Y., Chen, Y., Yang, X., & Zhang, G. (2022). Group risk assessment in failure
mode and effects analysis using a hybrid probabilistic hesitant fuzzy linguistic MCDM
method. Expert Systems with Applications, 188: 116013. DOI:

10.1016/j.eswa.2021.116013
Xie, X., Sun, F., Yang, X., Yang, Z., Gao, J., Ou, W., & Cui, B. (2021). Explore user neighbor-
hood for real-time e-commerce recommendation. In 2021 IEEE 37th International Con-
ference on Data Engineering (ICDE), 2464-2475. IEEE. DOI:

10.1109/ICDE51399.2021.00279
Zhang, Q., Zhu, Y., Ma, R., Du, C., Du, S., Shao, K., & Li, Q. (2022). Prediction Method of
TBM Tunneling Parameters Based on PSO-Bi-LSTM Model. Frontiers in Earth Science,
10: 854807. DOI.org/10.3389/feart.2022.854807
Zhang, X., Liu, H., Chen, X., Zhong, J., & Wang, D. (2020). A novel hybrid deep recommen-
dation system to differentiate user’s preference and item’s attractiveness. Information
Sciences, 519: 306-316. DOI.org/10.1016/j.ins.2020.01.044
Zheng, J., Li, Q., & Liao, J. (2021). Heterogeneous type-specific entity representation learning
for recommendations in e-commerce network. Information Processing & Management,
58(5):102629. DOI.org/10.1016/j.ipm.2021.102629
Zhou, K., & Zha, H. (2012). Learning binary codes for collaborative filtering. In Proceedings
of the 18th ACM SIGKDD international conference on Knowledge discovery and data
mining, 498-506. DOI.org/10.1145/2339530.2339611
Zhou, L. (2020). Product advertising recommendation in e-commerce based on deep learning
and distributed expression. Electronic Commerce Research, 20(2): 321-342. DOI:

10.1007/s10660-020-09411-6
 27

Zhu, B., Yang, C., Yu, C., & An, F. (2018). Product image recognition based on deep learning.
Journal of Computer-Aided Design & Computer Graphics, 30(9): 1778. DOI:

10.3724/SP.J.1089.2018.16849