Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Applied Mathematics and Nonlinear Sciences

https://www.sciendo.com

An Accuracy Study of Personalized Recommendation System for E-commerce Based

on Big Data Analysis

Hua Zhang1,†
1. School of Airport Economics and Management, Beijing Institute of Economics and Management,
Beijing, 100102, China.

Submission Info

Communicated by Z. Sabir
Received March 28, 2024
Accepted June 21, 2024
Available online August 5, 2024

Abstract
E-commerce, as an emerging value chain model for the global economy, has greatly promoted development, while the
impact of digitalization on traditional publishing enterprises is increasingly evident. In this paper, we propose a TextRank
keyword extraction algorithm based on comprehensive weights, which extracts and assigns keywords that identify user
information, behavior, and product characteristics. We then output a keyword weight table for user information, user
behavior, and product keywords. Finally, utilizing an optimized collaborative filtering recommendation algorithm, we
establish a recommendation model between the user-commodity matrix to build an e-commerce personalized
recommendation system that provides users with more accurate customized recommendations. The practical application
of the designed personalized recommendation system is examined to evaluate its accuracy. The MAE of this algorithm is
smaller than that of user-based (0.8915, 0.9470) or item-based (0.8873, 0.9327) collaborative filtering algorithms,
indicating that the improved collaborative filtering algorithm effectively enhances system recommendation accuracy. The
direct effect value of recommendation strength is 0.344, with an indirect effect value of 0.018, leading to the highest
overall effect value. This study provides users with convenient and attentive services, significantly enhances user
experience quality, and generates substantial profits for the e-commerce platform.

Keywords: SW-TextRank algorithm; Collaborative filtering; Recommender system; e-commerce.

AMS 2010 codes: 62-07

†Corresponding author.
Email address: [email protected] ISSN 2444-8656
https://doi.org/10.2478/amns-2024-1923
© 2024 Hua Zhang, published by Sciendo.
This work is licensed under the Creative Commons Attribution alone 4.0 License.
 2 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

1 Introduction

E-commerce personalized recommendation system, an advanced intelligent commerce system used

by e-commerce websites, provides customized product recommendations and information services
for customers shopping on the website through data mining technology [1-3]. In the information-
flooded network environment, e-commerce websites will face the test of how to effectively provide
users with the information they need, and the e-commerce personalized recommender system is the
e-commerce website customer one-to-one guide to help customers realize their personalized needs
[4-7]. E-commerce personalized recommender systems are widely used in e-commerce websites and
play an important role in providing fast and accurate product recommendations for website customers,
promoting customer purchases, providing personalized services, and improving the cross-selling
ability of the website and other aspects play an increasingly important role [8-11]. A recommender
system is an application that recommends products for target users based on their historical behavior
and preference information, and it is also an important research field of data mining, which is widely
used in the field of e-commerce. The most important module in the recommender system is the
recommendation algorithm, which can be roughly divided into collaborative filtering
recommendation, content-based recommendation, knowledge-based recommendation, hybrid
recommendation algorithms, etc. Based on the big data analysis of e-commerce personalized
recommender systems, it can improve the prediction accuracy as well as the higher computational
complexity brought by a large amount of data [12-15]. Therefore, many new solutions have been
generated to improve the similarity calculation method to enhance the prediction accuracy and so on.

Literature [16] took the flavors_of_Cacao dataset as the object of study and used neural network
algorithms such as decision trees to study and validate the dataset generated by e-commerce
recommender systems to minimize problems such as cold start and grey goat users in the
recommender system, in order to improve the performance of e-commerce recommender systems.
Literature [17] examined the collaborative filtering algorithm embedded in a typical online clothing
recommendation method, analyzed and evaluated the number of user purchases and the frequency of
inverse users in terms of precision, recall, and diversity, and the results showed that the method is an
effective way to improve the recommendation performance and promote the booming development
of online clothing e-commerce. Literature [18] studied the advantages of e-commerce operations
based on big data analysis. E-suppliers need big data analytics to understand consumer behavior, and
recommendation systems generated by big data technology can personalize recommendations to
consumers and improve the e-shopping experience for both consumers and suppliers. Literature [19]
proposed an e-commerce personalized recommendation algorithm for the fusion of particle-level
compressed perception and e-commerce customer trust relationship by using particle distance, and
the results show that the algorithm can solve the sparse problem of perception data and enhance the
accuracy of e-commerce personalized recommendation. Literature [20] prepares different
recommendation paths for active and inactive users, proposes a differentiated recommendation
framework, collects information about goods and users, strengthens fine-grained classification of
goods, discovers the potential relationship between goods and users, and accurately personalizes
recommendations for users with different characteristics. Literature [21] proposes an AFETLER
model combined with adaptive fusion technology, which integrates and analyzes recommendations
from multiple channels based on the user’s click-through rate, historical preferences, etc., to improve
the accuracy and overall recommendation quality of e-commerce personalized recommendation
systems.

The main process of the e-commerce personalized recommendation model in this paper is to extract
the keywords containing user information characteristics, user behavioral characteristics, and product
characteristics through the SW-TextRank algorithm and use the information of word properties, word
length, and word location to comprehensively determine the importance of the keywords. The
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 3

collaborative filtering recommendation algorithm is optimized based on the data organization model
of a hierarchical tree, and the recommendation model between the user and commodity matrix is
established. According to the structure and principle of the e-commerce personalized
recommendation system mentioned in the previous section, the e-commerce personalized
recommendation system is designed based on computer programming language. Combined with the
experimental data, the e-commerce personalized recommendation system is analyzed, and at the same
time, in order to further improve the verification of the e-commerce customized recommendation
system, the survey statistical method and structural equation model are used to explore the application
of the e-commerce personalized recommendation system.

2 Collaborative Filtering Based Personalized Recommender System for E-Commerce

The main process of the e-commerce personalized recommendation model in this paper is to extract
keywords that can identify the characteristics of user information, user behavior, and commodity
characteristics through the keyword extraction algorithm and use the information of word properties,
word length, and word location to determine the importance of the keywords, and then propose the
TextRank keyword extraction algorithm based on the integrated weights to assign weights to the
keywords, and then output User information keyword weight table, user behavior keyword weight
table and commodity keyword weight table, and finally use the collaborative filtering
recommendation algorithm based on similarity optimization to establish a recommendation model
between the user-commodity matrix to provide users with more accurate personalized recommended
content.

2.1 Personalized Recommendation Model Feature Extraction

In order to better realize feature extraction, this paper designs and implements a comprehensive
weight-based TextRank keyword extraction algorithm (SW-TextRank algorithm) to extract keyword
information from the content text of the user-goods matrix source data, streamline the user-goods
matrix source dataset, reduce the spatial scale of the source data, and reduce the useless data
information [22]. The core idea of the TextRank algorithm based on comprehensive weight is to
combine the display position Pos of information keywords, the length Len of information words, the
lexical Cla of information words, and the value of TF-IDF to assign a comprehensive weight to the
information keywords, and then use the extensive weight of these four indexes as the original weights
of TextRank algorithm for the keyword extraction. Keywords can reduce the irrelevant information
in the system’s user-commodity matrix source data, thus improving the data effectiveness of the
personalized recommendation model. [23]

2.1.1 Detailed design of composite weights

Comprehensive weight is a judgment made by analyzing the position of keywords in the text, their
length, linguistic properties, and TF-IDF value. Among them, the position of information keywords
appearing in the text Pos mainly includes three types: in the title, at the beginning of the paragraph,
and at the end of the paragraph. By analyzing the keywords in different positions it can reflect the
importance and criticality of the word in the text, and the position of keywords appearing in the text
Pos marking methods are as follows:

1, Information words in the title:


Posi = 2, The information word is at the beginning of the paragraph; (1)
3, Information words at the end of paragraphs;

 4 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Where Posi indicates the location identification information of information word i , the length of
the information word can indicate the key information of the information, according to the
classification standard of the information word in the knowledge network, the length of the
information word mainly consists of five levels, the information word length level identification are
as follows:

Leni = {Leni = 1, Leni = 2, Leni = 3,3  Leni  6, Leni  6} (2)

Where Leni indicates the length information of the information word i , information lexicality is
also one of the important indicators of the measurement of keywords. This paper focuses on the
information words are mainly divided into four kinds: nouns, adjectives, gerunds, and abbreviations
of four. Information lexicality is identified as follows:

Clai = {Clai = 1, Clai = 2, Clai = 3, Clai = 4} (3)

Where Clai represents the linguistic information of information word i , assuming that the
integrated weight of textual information word i at point v is W ( vi ) , the weight of TF-IDF is
W1 ( i ) , the positional weight of the information word is W2 ( i ) , the display length of the information
is W3 ( i ) , and the lexicality of the information is W4 ( i ) , then the integrated weight W (vi ) is
calculated by the following formula:

W ( i ) = (TF − IDE )*W1 ( i ) + Leni*W2 ( i ) + Clai*W3 ( i ) + Posi*W4 ( i ) (4)

Among them, the comprehensive weight W (vi ) is assigned by way of the G1 assignment method,
which has been elaborated in Chapter 2, and the expression of the assigned weight of each keyword
is as follows:

K i −1
ri = , (i = n, n − 1,..., 2) (5)
Ki

Where Ki −1 the l −1 nd importance of the text information, ri for the keyword i assignment
weight, through the assignment weight table can be queried for the assignment weight of each
keyword, the comprehensive weight assignment calculation formula is as follows:

1
W ( i ) = n n
1 +  ri (6)
i = 2 k =i

Then, substitute Eq. (6) into Eq. (4) to get the integrated weight of each keyword. After the integrated
weight calculation, the system converts the user-commodity matrix source data into a graph of user
keywords, commodity keywords, keywords of users’ behavior towards commodities, and the
corresponding weights of each keyword, and the graph is used as an input to the TextRank algorithm
to obtain all user-commodity keyword weights. Relationships.
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 5

2.1.2 Detailed Design of SW-TextRank Algorithm

Figure 1 shows the flowchart of the SW-TextRank keyword extraction algorithm based on SW-
TextRank. Because the original weights of the TextRank algorithm are all 1, which can not better
reflect the importance of keywords with different problems, this paper proposes a TextRank algorithm
based on the comprehensive weight. The core idea of the algorithm is to obtain the keywords
according to the four key weight indicators in the design of the comprehensive weights combined
with the G1 assignment method. The core idea of the algorithm is to obtain the comprehensive weight
of the keywords based on the four key weight indicators in the comprehensive weight design,
combined with the G1 assignment method, and then use the comprehensive weight of the keywords
as the initial weight of the TextRank algorithm and the keyword extraction algorithm. The TextRank
algorithm, based on comprehensive weights, assigns different weights to each keyword in the text,
which effectively improves the accuracy of the extracted keyword features. The formula for the SW-
TextRank algorithm is as follows:

W ji
S ( i ) = (1 − d )*W ( i )*  S ( j )

i In ( j) 
 k Out (
W jk (7)
j)

Where S ( vi ) represents the weight representation of point vi , all the keyword information can be
derived through the iterative calculation of the TextRank algorithm. The main process of the SW-
TextRank algorithm has the following steps:

1) Pre-processing of target text information: the target text information in this paper is the content
of the user information table, product information table, and user behavior information table.
The main step of pre-processing is to annotate the information location, information word
lexicality, information word length, etc., in these target texts, to filter some useless
information words, and to reduce the spatial size of the information candidate set.

2) Statistics of the return value of the four key indicators: for each keyword, to calculate the TF-
IDF value of the word, the record of the word is located in the location of information Posi .
Lexical Clai , the length of the keyword Leni , and other information, such as the information
contained in the text of the paragraph at the beginning of the “toothbrush”, which is the return
value of: Posi = 2, Clai = 1, Leni = 2 .

3) Calculate the comprehensive weights of all keywords according to the extensive weight
calculation formula and the G1 assignment method. The comprehensive weight calculation
formula is shown in (6) and will not be repeated here.

4) Construct the graph model using the TextRank algorithm to calculate the final keyword
weights: the keywords as the vertices of the graph, the comprehensive weights of the keywords
as the edges of the graph, construct the graph model, the final weights using expressions 3-7,
iterative calculations up to the point of convergence.

5) Sort the converged weights, sort the keywords by importance, filter the keywords using the
TOP-N method, replace the data in the source dataset of the user-commodity matrix, and
quantize the text data according to certain rules, for example, the gender keyword can be used
to indicate 1 for male: -1 for female.
 6 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Figure 1. Based on the SW-TextRank keyword extraction algorithm flowchart

2.2 Collaborative filtering recommendation algorithm and its improvement

One of the important algorithms for personalized recommendation is the collaborative overfiltering
algorithm [24]. In this section, the collaborative filtering algorithm is studied, its advantages and
disadvantages are analyzed, and then an improvement algorithm for the collaborative algorithm is
proposed based on the conceptual tree model to address the shortcomings of the collaborative filtering
algorithm.

2.2.1 Collaborative Filtering Recommendation Algorithm

The popularity of systematic filtering recommendation algorithms in e-commerce is due to their

ability to provide personalized e-commerce product recommendations. Currently, there are two types
of collaborative algorithms: item-based and user-based.

1) User-based collaborative filtering recommendation algorithms

Currently, the most successful personalized recommendation algorithm is the user-based

collaborative algorithm, which implements the recommendation through the rating data of the
immediate neighbors, the list of target users generated according to the target user, the degree
of recommendation of the personalized view similar to some of the rating system [25]. The
algorithm is based on the assumption that if the user’s interest in a group of items is more
similar, then their interest in other items is also more similar. The general process of a user-
based collaborative filtering recommendation algorithm is as follows:

(1) Data representation

 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 7

( )
The algorithm describes the user rating data as a m * n -order matrix A m, n

representation, with m representing the set of users and n representing the set of items,
and the i th row and j th column elements of the matrix R representing the rating
i, j

scores of the items j using i . The ratings are represented by binary 1s and 0s to
indicate the user’s preference (like/dislike) or purchase result (success/failed), or they can
be graded to describe the user’s preference for the item (in the database MovieLens ,
which is used later in the implementation and testing, the user’s preference for the movie
is represented by a number between the integer 5 and the integer 1.

(2) Nearest neighbor formation.

A “nearest neighbor” is a so-called shortest distance set that a recommendation algorithm

sets in order to find the most similar set. In this algorithm, for each customer u, a set of
“neighbors” N = {N , N , , N }, uN is generated in descending order of similarity.
1 2 k

1 s ( i)
The similarity algorithm from N to N can be represented by company sim u, N .

There are various methods of measuring similarity between customers, the main three of which are
cosine similarity, modified cosine similarity, and correlation similarity:

Cosine similarity: the ratings of user ui and user ui in n -dimensional space are represented as
vectors ui and u j to produce a model of the behavioral characteristics of customers ui and u j ,
then the similarity sim(ui , u j ) between customer ui and customer u j is calculated as shown in
Equation (8) as:

ui u j
sim(ui , u j ) = (8)
ui * u j

In Equation (8), the numerator is the inner product of the two customer rating vectors, and the
denominator is the product of the two customer vector modes.

Modification of cosine similarity: since different ratings are not the same, it will cause the accuracy
of prediction to be different. The modified cosine similarity is to take this problem into account for
improvement. The main practice is to subtract the average user rating of the project.

Correlation similarity: that is, the parameters of user evaluation are measured, and their intersection
I  I , is taken, then the similarity w ( a, i ) between user i and user j is can be calculated by
equation (9):

 (raj - ra )(rij - ri )
w ( a, i ) =
jI a  I i
(9)

jI a  Ii
(raj - ra ) 2 
jI a  Ii
(rij - ri ) 2

Where ra and ri are the average scores of users a and i respectively. Depending on the size of
w ( a, i ) , N, similar users are selected as “nearest neighbors” of the target user.
 8 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Generating a recommendation is to recommend some information needed by the user from the
“nearest neighbors” obtained from the correlation similarity analyzed above. The predicted score Pu , j
of user u for item j can be calculated by equation (10):

N
Pu , j = ra + k  w(u , i )(rij - ri ) (10)
i =1

This algorithm can calculate the information needed by the user. That is, the first number of items
(Top − N ) with the highest measurement score is recommended to the target user as a
recommendation result.

2) Item-based collaborative filtering recommendation algorithm

The algorithm filters the resources and filters out some things that are of interest to the user to
the customer [26]. In this algorithm, the set of users is an ensemble of users who have
participated in the evaluation of item i and item j together. The similarity between item
i and item j is obtained by similarity calculation sim ( i, j ) . The following is the formula
for calculating the relevant similarity as shown in (11):

 (R
uU u
u ,i − Ru )( Ru , j − Ru )
sim(i, j ) = (11)
 (R
uU i , j
u ,i − Ru ) 2  (R
uU ij
u, j − Ru ) 2

Where Ru denotes the average rating of user u on the item. Using Pu , p to denote the weighted
resulting predicted ratings, then the predicted ratings Pu , p of user u for item p can be shown by
equation (12):

 sim
nN p
p ,n  Ru ,n
Pu , p = (12)

nN p
sim p ,n

Finally, the top number of items (Top − N ) with the highest predicted ratings are recommended to
the target users.

2.2.2 Improvement of collaborative filtering algorithm

This paper uses a hierarchical tree-based data organization model to improve the existing
collaborative filtering recommendation algorithm, to comprehensively improve the accuracy of the
e-commerce recommendation system, and to effectively complete the recommendation strategy of
the e-commerce system recommendation system.

In order to overcome the above shortcomings, the performance of the algorithm has to be improved,
and the performance is changed by representing the interest and preference model of user ui by the
conceptual hierarchical tree-based user item type scoring vector Vi . The vector representation is
shown in equation (13):
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 9

Vi = ( i1 , i 2 , , ik ,  i|C| ) (13)

Which Vi is the user i ; for item types, C is the total number of nodes in the hierarchical tree T ,
and Ci  C the evaluation score.

In order to enable the user interest model to reflect the distribution of user interests within a unified
scale and to ensure the consistency of user rating criteria, this method establishes a unified score scale
for user ratings of all item types. The total score S is the sum of user ui ratings for each item
category, which is calculated by Equation (14):


|c|
k =1 ik
v =S (14)

Let the evaluation score of user ui be set Ri = ri1 , ri 2 , , rij , rin  , and the evaluation score of user
ui for item i j be rij , i j which can be categorized under f (i j ) item categories in the subset
f (i )  C of item categories. The process of assigning S to all item categories in the set C of item
categories is as follows:

1) In proportion to the ratio of score i j to the total score of user ui ’s evaluated score, the
corresponding obtained score of the score item i j will be obtained from the sum of ratings
S. The score will then be distributed equally to each category of the various item types to
which i j belongs to obtain an initial score S (c jk ) of the item type c jk  f (i j ) to which
the user ui ‘s evaluated item i j belongs to which can be obtained by the formula (15):

rij
S (c jk ) = S  n
f (i j ) •  rij
(15)
j =1

2) Let ( p0 , p1 , pt , pk ) denote a path from the top node po to the bottom node pk = c jk
of a concept tree. The initial score S ( c jk ) obtained from the score of item type c jk
corresponding to item i j is assigned to each node of path ( p0 , p1 , pt , pk ) according to
a certain allocation rule, and the score S ( pi ) shared by each level of nodes pi in the path
is obtained by Eq. (14) and Eq. (15):
k

 s ( p ) = S (c
t =0
t jk ) (16)

s ( pt +1 )
s ( pt ) = (17)
b( pt +1 + 1)

Where b ( pi ) denotes the number of brother nodes of node pi .

 10 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

3) Add the computed score s ( pi ) to the component vu in the scoring vector Vi

corresponding to the item kind ci represented by node pi .

4) Repeat step 3) to compute the scores of the vector components corresponding to each node on
the path where the other item kinds to which item i j belongs are located.

5) Repeat steps 1) to 4) to convert the user ui ’s evaluation scores for the remaining rated items
into evaluation scores for the corresponding item categories, thus completing the creation of
the interest model for user ui .

2.3 Front-end design of e-commerce personalized recommendation system

Page display is the interface between the e-commerce personalized recommendation system and the
user. The design of a reasonable layout of the customized recommendation module to meet the user’s
page browsing needs is the front-end design of the personalized recommendation system is more
critical. The following will elaborate on the e-commerce personalized recommendation system of the
front-end design of the main modules.

2.3.1 Home page recommendation module design

For tourists and newly registered users, the home page personalized recommendations should be
obvious and easy to make the user interested, so in the home page module page design, the
recommended products should be presented on the first screen of the user’s page, at the same time,
the size of the product image should be slightly larger than the size of the image of the other modules,
you can also use the picture rotation function to recommend to the user. At the same time, the product
information should include the market price and the site price, which can stimulate the consumer’s
psychology to open the page and browse the product details.

2.3.2 Browse Pages - Similar Product Recommendations

When a user browses a product page, a series of label lists are generated by the name, description,
and keywords of this product, and this page, and then an algorithm is utilized to query related similar
products to recommend to the user. Similar products are displayed in the form of a list, providing
product information, mostly including product images, names, prices, etc.

2.3.3 Purchasing Products-Related Product Recommendations

When the user has already purchased a product, this time recommendation module shows the user the
related products of the product. That is, the user who has purchased this product has also purchased
which product. The page mainly displays the name of the product, picture, clicks, sales, and so on.

2.4 E-commerce personalized recommendation system design

The design of the e-commerce personalized recommendation system mainly includes the design of
the overall system architecture, the design of the solution to the cold-start problem, and the
construction of the user behavior model based on hierarchical vector controls.
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 11

2.4.1 Overall architecture

The E-commerce personalized recommendation system is mainly for three types of users: non-
registered users (visitors), newly registered users, and old users, different users need to take different
strategies to achieve the optimal personalized recommendation. The E-commerce personalized
recommendation system’s overall architecture is shown in Figure 2. Non-registered users come to the
site according to its source of recommendation, whether it is through the search engine by entering
the site’s name or in the browser to enter the URL to the home page, then in the items recommended
module to the user recommended site popular products, if the user is searched through the keywords,
then the use of solr full-text search, retrieval of products that contain the keywords in the list of
recommended to show to the user. For newly registered users, they can obtain their label from their
registration information and according to their tags, recommend products related to this tag
classification to them. For older users, their previous browsing, collection, purchase, evaluation, and
other information can be used to provide personalized recommendations. These three
recommendation modes can be in the recommendation list for the user to recommend the results of
feedback so as to further optimize the results of the recommendation for the more appropriate
products recommended to the user.

Figure 2. Electronic business personalized recommendation system architecture

2.4.2 Cold start problems

For users who have not yet registered on the website, on the one hand, you can give the most
appropriate recommendations by checking their source information, and on the other hand, you can
recommend the popular products on the website to the users. Newly added items can be added to the
popular recommendations for some time to observe the number of clicks and other information before
deciding on the next action. To address the cold start issue of registered users, this system utilizes a
combination of labeling recommendations and hot lists. Users can tag themselves with their favorite
labels when registering, the registration page gives the classification labels of the products on the
website, and users can mark the classifications they are interested in. When users register successfully,
the system will store this information in the database for subsequent recommendation.

Initially, the system will be based on the user’s label, and the hot list of relevant recommendations
first determines whether the user’s label can be found in the hot list of the corresponding product. If
it exists, the hot list of the corresponding product is directly recommended to the user. If it does not
exist, then the use of Solr full-text search in the full product database index query, to recommend to
the user the product with the highest click-through rate. With the increase in user behavior, the original
 12 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

labeling system is not enough to reflect the real needs of the user at this time, according to the user
behavioral preferences and then combined with the labeling system to build a user model and then
personalized recommendations.

2.4.3 Behavioral preferences of users

In this paper, the improved collaborative filtering recommendation algorithm proposed above is used
to construct a user behavioral feature model, where each user can have multiple different preferences,
and each preference can correspond to various products for which the user generates behavior. The
user’s behavioral preferences will change over time. When the user does not have a behavior
corresponding to a certain behavioral preference for a longer period, this value is deleted from the
behavioral preference list. If there is a new user behavioral preference, the first step is to determine
whether the number of user behavioral preferences reaches a certain threshold. Then the weight of
the new user behavioral preference is calculated. If the number of user behavioral preferences has not
yet exceeded the threshold, then the data is direct. If the number of user behavioral preferences has
not yet exceeded this threshold, add this data to the user’s behavioral preference data, and if the
number of user behavioral preferences exceeds this threshold, replace this behavioral preference with
the behavioral preference with the smallest weight value saved in the database, and complete the
update of the user behavioral preference model.

2.4.4 Personalized Recommendation Generation

On the basis of the user behavioral preference model and product attribute model, the similarity
between users and products is firstly calculated using the Person correlation algorithm, then the
nearest neighbors and initial recommendation results are obtained, and finally, the final personalized
recommendation results are obtained using improved filtering based collaborative recommendation
algorithm.

3 Accuracy and Application of Personalized Recommendation System for E-commerce

3.1 Accuracy analysis of e-commerce personalized recommendation system

3.1.1 Experimental data and environment

The hardware platform used for the experiments in this paper is an Intel Core4 CPU+4G RAM
computer. The running environment is Windows 10+Eclipse+Mysql, and the programming language
is JAVA. The experimental data comes from the Movie Lens dataset provided by the Group Lens
Research Program at the University of Minnesota. It is the user rating data for movies, and there are
5 ratings from 1 to 5. The main structure is made up of 6 structures: core_genre, core_movie,
core_movie_genre, core_occipation, core_rating, core_user. Movie Lens dataset Movies have 15
categories with a total of 80,000 ratings, which are compiled by 937 users rating 1,641 movies. This
dataset can better reflect the differences in attributes between items, and it is also commonly used to
validate personalized recommendation algorithms. The dataset’s sparsity is 0.87, which is very sparse.
During the experiment, by default, this dataset is randomly divided into 2 parts: 75% is the training
set, and the other 25% is the test set.
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 13

3.1.2 Experimental program

The purpose of the experiment is to verify the feasibility of the improved collaborative filtering
algorithm proposed in this paper, firstly, to confirm whether the improvement of the formula for
similarity calculation is feasible, and secondly, it is also necessary to verify whether the improved
algorithm improves the accuracy of the recommendation. The formula will be used to measure the
accuracy of the recommendation results for this experimental purpose. The specific implementation
scheme of the experiment is to take the training set data as known, calculate a total of 20,000 predicted
scores for the items present in the test set respectively, and then calculate the MAE between the
predicted scores and the real scores according to the formula, and compare the accuracy between the
algorithms through this MAE. Where the size of the neighbor set n is taken from 10 to 100 in
increments of 10. The experiment consists of two small experiments:

Experiment 1: The purpose of this experiment is to determine whether the improvement of the
formula is feasible. The experiment utilizes cosine similarity, person correlation, and enhanced
similarity between users.

Experiment 2: The purpose of this experiment is to verify whether the improved collaborative filtering
algorithm enhances the accuracy of the recommendation results. The experiment uses the improved
collaborative filtering recommendation algorithm proposed in this paper, user-based collaborative
filtering algorithm, and item-based collaborative filtering algorithm to compute the predicted scores,
respectively.

Experiment 3: The purpose of this experiment is to further verify the feasibility of the improved
collaborative filtering algorithm. The size of the training set is modified on the basis of Experiment
II, and the proportion occupied by the training set is increased from 0.2 to 0.9 with an increment of
0.1. The neighbor set’s size remains constant at 100.

3.1.3 Analysis of experimental results

Through this MAE, to compare the accuracy between the algorithms, the smaller the MAE value, the
higher the accuracy. The results of the analysis of Experiment 1 are shown in Figure 3. From the
experimental results, it is easy to see that the improved similarity MAE value proposed in this paper
is smaller and better compared to the cosine similarity (0.8930), and Person similarity (0.8752) is
lower, the improved similarity formula proposed in this paper can improve the accuracy of the
recommended algorithms, and the improvement of the similarity calculation for the similarity
calculation is feasible.

Figure 3. Experimental 1 analysis results

The results of Experiment II analysis are shown in Fig. 4, while the results of Experiment III analysis
are shown in Fig. 5. From the results of Experiment II and Experiment III, it is easy to see that with
the increase of the number of neighbor sets, the MAE is decreasing, which is because the rise of the
number of neighbor sets leads to the increase of the accuracy of the recommendation results.
 14 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Meanwhile, with the rise in the training set, MAE also shows a decreasing trend. This is because the
increase in the original rating data will improve the accuracy of the recommendation results. In the
above 2 experiments, the improved collaborative filtering recommendation algorithm proposed in
this paper has a smaller MAE than the user-based (0.8915, 0.9470) or item-based (0.8873, 0.9327)
collaborative filtering algorithms on the whole (0.8774, 0.9228), (the former one is the MAE value
of Experiment 2 and the latter one is the MAE value of Experiment 3). We all know that the MAE
and the accuracy of the recommendation algorithm are inversely proportional to each other, which
shows that the improved collaborative filtering algorithm is effective in improving the accuracy of
the recommendation results and also fully proves the effectiveness of the enhanced collaborative
filtering algorithm proposed in this paper.

Overall, the improved collaborative filtering recommendation algorithm proposed in this paper is
significantly better than the traditional collaborative filtering recommendation algorithm in terms of
the accuracy of recommendation results. This is because the conventional collaborative filtering
algorithms, such as the calculation of similarity, do not take into account the differences between the
user’s points of interest and the number of common rating items caused by the similarity of the
calculation of the deviation is large, the proposed for the correlation between the items instead of the
differences between the user’s points of interest, and the difference in the number of common ratings
between the users into the calculation of user similarity, thus obtaining more accurate
recommendation results. As a result, recommendations can be more precise.

Figure 4. Experimental 2 analysis results Figure 5. Experimental 3 analysis results

3.2 E-commerce personalized recommendation system application analysis

3.2.1 Research Objects

The descriptive statistics of the official questionnaire sample are shown in Table 1. The total number
of valid sample sizes returned was 250. Among the valid samples, 100 were males, accounting for
40.00%, and 60.00% were females. In terms of age distribution, the samples obtained in this study
are mainly people under 35 years of age who have experience in online shopping, of which the number
of samples for 18-22 years of age is 99, accounting for 39.60%. The number of samples for 22-25
years of age is 102, accounting for 40.80%. The number of samples for 25-28 years of age is 39,
accounting for 45.60%. The number of samples for 28-35 years of age is 15, accounting for 4.00%.
From the distribution of samples, it can be seen that the age of online shopping consumers is mainly
concentrated in the age of 18-35 years old, which is related to the background and object of our
research. As the research in this paper is aimed at the online purchase intention supported by the
recommender system, we are mainly more likely to over-consume, and the crowd volume is relatively
large, easy to accept the emerging things for the main survey object, so the age distribution is
generally in the age of 18-35 years old. In terms of education, the distribution of education is mainly
concentrated on undergraduate and postgraduate students, with a small number of PhDs and
specialists. In terms of frequency of online shopping, a total of 185 people, accounting for 74.00%,
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 15

shopped online more frequently. In addition, 197 people used the personalized recommendation
system, accounting for 78.80%.

Table 1. Descriptive statistics of formal questionnaires

Sample statistics Type Frequency Percentage
Man 100 40.00%
Gender Female 150 60.00%
Total 250 100.00%
18-22 99 39.60%
22-25 102 40.80%
Age 28-28 39 15.60%
28-35 10 4.00%
Total 250 100.00%
Specialty 12 4.80%
Undergraduate 114 45.60%
Education degree Graduate student 108 43.20%
Doctor and above 16 6.40%
Total 250 100.00%
Often 185 74.00%
Shopping Frequency Occasionally 65 26.00%
Total 250 100.00%
Yes 197 78.80%
Whether to buy goods according to the recommendation system of
No 53 21.20%
the shopping website personalized recommendation system
Total 250 100.00%

3.2.2 Descriptive analysis

A1 is that the recommended information is organized in an orderly manner, A2 is that the

recommended information is organized in accordance with the norms, A3 is that the recommended
information is organized in a more rationalized way, B1 is that the recommended product information
of the recommendation system meets my preferences and needs, B2 is that the recommended product
information of the recommendation system is very complete, C1 is that the personalized
recommendation system recommends professional information, C2 is that the information provided
by the customized recommendation system is true C3 is that the information recommended by the
personalized recommendation system is not biased, C4 is that the customized recommendation system
does not intentionally provide false information, D1 is that the information recommended by the
customized recommendation system is too much, D2 is that the personalized recommendation system
recommends information too often, D3 is that there is no time or energy to read the information
recommended by the customized recommendation system, E1 is that the recommended information
is vivid, E2 is that the recommended information is accompanied by audio or video descriptions, E3
is that the recommended information contains comments from other consumers, F1 is that the website
is well known, F2 is that the website has a high reputation, F3 is that the website is an authority in its
field, F4 is that the website is reliable, F5 is that the website is trustworthy, G1 is that I am willing to
trust other people even though I don’t know them very well, G2 is that I tend to trust other people
unless there is a reason for distrust, G3 is that even if I don’t know them very well, G3 is that even if
I don’t know them very well, I tend to trust them. Trust others, G3 is I would trust others even if I
don’t know them very well, H1 is I think the recommended products or services are of good quality,
H2 is I guess the recommender system makes recommendations based on real situations, H3 is I
believe that the recommendation behavior is not purely for the sake of profit, H4 is I would like to
think that the recommendation behavior is well-intentioned, H5 is I believe that the recommended
 16 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

products or services can satisfy my needs, H6 is I believe that the recommender system is rich in
product knowledge, I for I think there is no better choice than buying the recommended products, I2
for I have a strong willingness to purchase the recommended products and services, and I3 for I am
willing to browse the product information pushed by the personalized recommender system
frequently. The questionnaire data was analyzed descriptively using SPSS 16.0 analysis software. The
sample data of mean and standard deviation, skewness and kurtosis, standard error of skewness and
standard error of kurtosis of the research data were examined to see if they conformed to normal
distribution, and the results of the descriptive analysis are shown in Table 2. From the table, it can be
seen that the absolute value of the kurtosis and skewness of the measurement index data obtained
from this questionnaire is less than 1. It shows that the data obtained from this survey is more
consistent with the requirements of a normal distribution.

Table 2. Descriptive analysis results

Symbol Mean SD Degree of bias Degree of bias deviation Kurtosis Kurtosis deviation
A1 3.266 0.985 -0.143 0.133 -0.242 0.229
A2 3.214 1.06 -0.355 0.133 -0.542 0.229
A3 3.439 1.06 -0.322 0.133 0.212 0.229
B1 3.396 0.842 -0.344 0.133 0.429 0.229
B2 3.037 0.925 -0.569 0.133 0.374 0.229
C1 3.605 0.912 -0.428 0.133 -0.291 0.229
C2 3.38 0.873 -0.432 0.133 -0.171 0.229
C3 3.435 0.841 -0.136 0.133 0.233 0.229
C4 3.488 0.877 -0.158 0.133 -0.219 0.229
D1 3.533 1.1 -0.333 0.133 -0.19 0.229
D2 3.511 0.902 -0.422 0.133 -0.14 0.229
D3 3.113 0.996 -0.323 0.133 -0.683 0.229
E1 2.788 0.996 0.177 0.133 -0.515 0.229
E2 3.404 0.878 -0.147 0.133 -0.364 0.229
E3 3.545 0.946 -0.246 0.133 0.318 0.229
F1 3.506 0.85 -0.282 0.133 -0.334 0.229
F2 3.647 0.889 -0.066 0.133 -0.297 0.229
F3 3.445 1.04 -0.074 0.133 0.305 0.229
F4 3.705 0.945 -0.016 0.133 -0.337 0.229
F5 3.392 0.91 -0.313 0.133 0.535 0.229
G1 3.709 0.821 -0.431 0.133 0.669 0.229
G2 3.036 0.886 0.244 0.133 0.263 0.229
G3 3.548 0.971 0.181 0.133 -0.377 0.229
H1 3.532 0.883 -0.144 0.133 0.321 0.229
H2 3.233 1.053 -0.268 0.133 -0.327 0.229
H3 3.425 1.021 -0.077 0.133 0.329 0.229
H4 3.428 0.825 -0.146 0.133 0.241 0.229
H5 3.6 0.911 -0.433 0.133 -0.521 0.229
H6 3.673 0.877 -0.076 0.133 -0.545 0.229
I1 3.529 0.888 -0.367 0.133 0.449 0.229
I2 2.559 0.993 0.193 0.133 -0.323 0.229
I3 3.431 0.838 -0.146 0.133 -0.176 0.229
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 17

3.2.3 Analysis of the application mechanism of the system on the increase of purchase
intention

This paper utilizes structural equation modeling to analyze the interrelationships among variables.
This method of analysis has the common point of both factor analysis and path analysis, effectively
analyzing the internal influence mechanism and a path between variables, and this method is an
emerging research method for social science research in academia. Two statistical software tools,
SPSS16.0 and AMOS22.0, are used in this paper to analyze the questionnaire data. The structural
equation modeling of the system and purchase intention enhancement is shown in Figure 6, the results
of the model path coefficients are shown in Table 3, and the overall effect values are shown in Table
4. The synthesis of Figure 6, Table 3, and Table 4 shows that, among them, the path coefficient
between consumer trust and purchase intention is 0.617, the C.R. value is 7.06, and there is a
significant correlation between them. The direct effect value of recommendation strength is 0.344,
while the indirect effect value is 0.018, and it ranks first in the overall effect value. Overall, the
greatest influence on consumer purchase intention is the strength of recommendations (0.362). In
order of “information reliability (0.326)”, “information organization (0.246)”, and “website image
(0.208)”. This indicates that in order to make the e-commerce personalized recommendation system
play a positive role, it is not only necessary to improve the reliability of the information recommended
by the e-commerce customized recommendation system but also necessary to improve the e-
commerce personalized recommendation system so that its recommended information arrangement
is more reasonable.

e24 e25 e26 e27 e28 e29 e21 e22 e23

e1 A1

e2 A2 Information H1 H2 H3 H4 H5 H6 G1 G2 G3
layout

e3 A3

e4 B1 Information Confidence
Trust
detail tendency
e5 B2

e6 C1

e7 C2 Information
reliability
e8 C3

e9 C4

e10 D1

Recommended
e11 D2
strength

e12 D3

e13 E1

Purchase
e14 E2 Visual cue
intention

e15 E3

e16 F1

I1 I2 I3
e17 F2

e18 F3 Website image

e30 e31 e32

e19 F4

e20 F5

Figure 6. Model of the structural equation of system and purchase intention

 18 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

Table 3. Model path coefficient result

Path Estimate C.R. P Significance
Trust→Information layout 0.508 7.392 *** Significance
Trust→Information detail 0.417 6.202 0.004 Significance
Trust→Information reliability 0.644 8.13 *** Significance
Trust→Recommended strength 0.207 3.237 0.218 No
Trust→Visual cue 0.263 4.904 0.111 No
Trust→Website image 0.513 6.928 *** Significance
Trust→Confidence tendency 0.619 7.863 *** Significance
Purchase intention→Trust 0.617 7.604 *** Significance
Purchase intention→Information layout 0.412 6.223 0.004 Significance
Purchase intention→Information detail 0.191 2.926 0.308 No
Purchase intention→Information reliability 0.479 6.727 0.004 Significance
Purchase intention→Recommended strength 0.127 1.424 0.431 No
Purchase intention→Visual cue 0.249 4.061 0.133 No
Purchase intention→Website image 0.347 5.38 0.006 Significance

Table 4. Overall effect

Variable name Direct effect Indirect effect Overall effect The overall effect is ranked
Information Layout 0.141 0.108 0.249 3
Information detail 0.031 0.047 0.078 6
Information reliability 0.18 0.146 0.326 2
Recommended strength 0.344 0.018 0.362 1
Visual cue 0.059 0.027 0.086 5
Website image 0.131 0.077 0.208 4
Trust 0.042 0.025 0.067 7

4 Conclusion

The impact of the new wave of globalization and digitization can be mitigated by the emerging
business model of e-commerce, which can improve the comparatively backward business situation
of the current publishing industry. In this paper, after enhancing the collaborative filtering algorithm,
we design an e-commerce recommendation system based on the collaborative filtering algorithm, and
in order to confirm the practicability of the developed system, we analyze the accuracy and
application of the e-commerce personalized recommendation system in view of big data. The research
findings are presented as follows:

1) The MAE of the improved collaborative filtering recommendation algorithm (0.8774, 0.9228)
in this paper is smaller compared to the user-based (0.8915, 0.9470) or item-based (0.8873,
0.9327) collaborative filtering algorithms, and it is found that the MAE decreases with the
increase of the number of neighbor sets. Overall, the improved collaborative filtering
recommendation algorithms are able to obtain more accurate recommendation results and
recommend accurate real-time information about e-commerce domains for users.
 An Accuracy Study of Personalized Recommendation System for E-commerce Based on Big Data Analysis 19

2) Through the analysis of the application of an e-commerce personalized recommendation

system in the field of users’ online purchase intention, it can be seen that the path coefficient
between consumer trust and purchase intention is 0.617, and there is a significant correlation
between the two. In addition, the greatest degree of influence on the consumer’s willingness
to buy is the recommendation strength of the personalized recommendation system (0.362),
indicating that the recommendation strength of an e-commerce personalized recommendation
system can be strengthened to better serve the user but also the development of e-commerce
industry has an important role in promoting.

Funding:

1) China Association of Higher Education project “Research on the Construction of Professional

Digital Transformation Models for Higher Vocational College based on digital technology”
(number: 23ZYJ0426).

2) The Beijing Digital Education Research Project “Digital Technology Empowerments Higher
Vocational College Professional Digital Transformation and innovation and Practical
Research” (number: BDEC2023619082).

References
[1] Patro, S. G. K., Mishra, B. K., & Panda, S. K. (2022). Hybrid action-allied recommender mechanism: an
unhackneyed attribute for ecommerce. ECS transactions(1), 107.
[2] Jannach, D., Ludewig, M., & Lerche, L. (2017). Session-based item recommendation in e-commerce.
User Modeling and User-Adapted Interaction.
[3] Xiaojun, L. (2017). An improved clustering-based collaborative filtering recommendation algorithm.
Cluster Computing, 20(2), 1281-1288.
[4] Márcio Guia, Silva, R. R., & Bernardino, J. (2019). A hybrid ontology-based recommendation system in
e-commerce. Algorithms, 12(11), 239.
[5] Cai, J. (2017). Research on personalized recommendation algorithm in e-commerce based on hybrid
algorithm. C e Ca, 42(2), 590-594.
[6] Choi, Y., & Kim, S. K. (2018). A recommendation system for repetitively purchasing items in e-
commerce based on collaborative filtering and association rules. Journal of Internet Technology, 19, 1691-
1698.
[7] Lv, J., Song, B., Guo, J., Du, X., & Guizani, M. (2019). Interest-related item similarity model based on
multimodal data for top-n recommendation. IEEE Access, 1-1.
[8] Zhang, J., Zhang, C., & Yu, H. (2018). Research on e-commerce intelligent service based on data mining.
MATEC Web of Conferences, 173(3), 03012.
[9] Li, Z., Amagata, D., Zhang, Y., Maekawa, T., Hara, T., & Yonekawa, K., et al. (2022). Hml4rec:
hierarchical meta-learning for cold-start recommendation in flash sale e-commerce. Knowledge-based
systems.
[10] Jiao, W., & Li, J. (2017). Information recommendation system based on collaborative filtering in cloud
computing. Boletin Tecnico/Technical Bulletin, 55(7), 98-104.
[11] Zeng, Y., Song, S., & Peng, W. (2022). Optimal add-on items recommendation service strength strategy
for e-commerce platform with full-reduction-promotion. RAIRO - Operations Research, 56(2), 1031-
1049.
[12] Zhang, N. (2018). A personalized recommendation algorithm based on text mining. International Journal
of Performability Engineering, 14(7), 1401-1410.
 20 Hua Zhang. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-20

[13] Wang, B., Ye, F., & Xu, J. (2018). A personalized recommendation algorithm based on the user’s implicit
feedback in e-commerce. Future Internet, 10(12), 117.
[14] Cui, Y. (2021). Intelligent recommendation system based on mathematical modeling in personalized data
mining. Mathematical Problems in Engineering, 2021(3), 1-11.
[15] Li, W., Zhou, X., Shimizu, S., Xin, M., Jiang, J., & Gao, H., et al. (2019). Personalization recommendation
algorithm based on trust correlation degree and matrix factorization. IEEE Access, 1-1.
[16] Kaur, B., & Rani, S. (2023). Are the customers receiving exact recommendations from the e-commerce
companies? towards the identification of gray sheep users using personality parameters. International
Journal of Performability Engineering(7), 19.
[17] Hu, Z. H., Li, X., Wei, C., & Zhou, H. L. (2019). Examining collaborative filtering algorithms for clothing
recommendation in e-commerce:. Textile Research Journal, 89(14), 2821-2835.
[18] Alrumiah, S. S., & Hadwan, M. (2021). Implementing big data analytics in e-commerce: vendor and
customer view. IEEE Access, PP(99), 1-1.
[19] Jia, Z. (2017). Cross-border e-commerce personalized filtering recommendation model based on particle-
level compression awareness. Boletin Tecnico/Technical Bulletin, 55(7), 136-143.
[20] Yan, C., Chen, Y., & Zhou, L. (2019). Differentiated fashion recommendation using knowledge graph
and data augmentation. IEEE Access, PP(99), 1-1.
[21] Bagga, V., Sugunan, S., Srivastava, A., Kumar, R., Gupta, A., & Kumar, D., et al. (2023). Adaptive fusion
and transfer learning for enhanced e –commerce recommendations. Procedia Computer Science, 229, 345-
356.
[22] Wenjun Liu,Yuyan Sun,Bao Yu,Hailan Wang,Qingcheng Peng,Mengshu Hou.. & Cheng
Liu.(2024).Automatic Text Summarization Method Based on Improved TextRank Algorithm and K-
Means Clustering.Knowledge-Based Systems111447-.
[23] Qiu, Dong & Zheng, Qin.(2021).Improving TextRank Algorithm for Automatic Keyword Extraction with
Tolerance Rough Set.International Journal of Fuzzy Systems(3),1-11.
[24] Ni Li & Yinshui Xia.(2024).Movie recommendation based on ALS collaborative filtering
recommendation algorithm with deep learning model.Entertainment Computing100715-.
[25] Sheng Jinfang, Liu Qingqing, Hou Zhengang & Wang Bin. (2023). A Collaborative Filtering
Recommendation Algorithm Based on Community Detection and Graph Neural Network.Neural
Processing Letters (6),7095-7112.
[26] Ying Yin.(2024).Research on the integration path of cultural creative industry and tourism industry based
on collaborative filtering recommendation algorithm.Applied Mathematics and Nonlinear Sciences(1),

About the Author

Hua Zhang (1978.12—), female, Liaocheng Shandong, Beihang University, Ph.D. graduate, School
of Airport Economics and Management, Beijing Institute of Economics and Management, associate
professor, main research direction: e-commerce, large number analysis.