Design and Application of Personalized Recommendation Algorithm

Model Based on E-Commerce Platform Data

Qi Ding

and Zhigang Zhu

City Institute, Dalian University of Technology, Dalian, Liaoning, China

Keywords: E-Commerce Platform Data, Personalized Recommendation, Collaborative Filtering Algorithm (CF), Data

Model, Python.

Abstract: The algorithm models of K-means, Item-CF and User-CF based on Python environment can well realize

various functions of personalized recommendation service system of e-commerce platform, and can formulate

different recommendation service strategies for different user groups, which can effectively solve the problem

of adaptation between e-commerce platform and users' needs and improve users' purchase efficiency and

experience. Therefore, this paper takes the running data of e-commerce platform as the research object, relies

on data processing class libraries such as Numpy and Pandas in Python environment, builds a personalized

recommendation engine, and forms a personalized recommendation service system adapted to the call of Web

Server through systematic encapsulation. Personalized recommendation service system will be between user

I/O interface and e-commerce platform, and adopt MVC technology framework as the core design, and design

API interfaces that can be called according to different application scenarios, so as to achieve a high degree

of integration between recommendation system and e-commerce platform, meet the recommendation service

strategy formulation requirements of e-commerce platform, and at the same time, it has good technical

expansion performance in improving personalized recommendation.

1 INTRODUCTION

E-commerce, which combines the application

advantages of computer science and technology,

network information technology and remote

communication technology, is a kind of commercial

activity centered on commodity exchange in the

virtual network environment. With all-round

intelligent terminal equipment, high-speed and

convenient network communication and safe and

transparent mobile payment function, consumers

enjoy the great convenience brought by constantly

updated and abundant shopping information, but also

have to suffer from information overload problems

such as excessive information bombing, information

expansion and information kidnapping. (Du, 2018)

On the one hand, the information overload

problem of E-commerce originates from the

explosive growth of E-commerce shopping

platforms; on the other hand, it is caused by the

increasingly perfect development of E-commerce

https://orcid.org/0000-0001-7956-4041

https://orcid.org/0000-0002-5782-4444

platforms and the increasing data and information

capacity. For the E-commerce platform itself, the

function deployment is becoming more and more

comprehensive, the whole system is becoming more

and more complex, and the data interaction behind it

is becoming more and more frequent. At the same

time, there are more and more stores and brands

carried by E-commerce platforms, and the number of

commodities has reached the order of one billion.

Diversified platform selection, rich search,

comparison and screening functions, massive

commodity information and a large amount of

marketing and promotion information can help users

to get the commodities they need, but it is often

counterproductive, prompting users to have negative

emotions such as anxiety and entanglement, which

affects the formulation of consumption decisions. In

addition, after more and more consumers have met

their basic needs, their consumption behaviors are

more mature and their consumption psychology is

more stable, so that their consumption needs are

310

Ding, Q. and Zhu, Z.

Design and Application of Personalized Recommendation Algorithm Model Based on E-Commerce Platform Data.

DOI: 10.5220/0012030000003620

In Proceedings of the 4th International Conference on Economic Management and Model Engineering (ICEMME 2022), pages 310-315

ISBN: 978-989-758-636-1

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

personalized and heterogeneous, which puts forward

higher requirements for the experience of E-

commerce platform, and has a great impact on the

formulation of consumption decisions. How to break

through the influence of E-commerce platform

information overload on consumers' shopping

decisions is a hot issue in the development of E-

commerce to improve consumer experience and

shopping efficiency. (Shang, 2021) In view of this,

this paper believes that E-commerce platform can

build a personalized and professional

recommendation service system. With the application

advantages of a large number of machine learning

algorithms in data processing, it can quickly realize

the design and deployment of recommendation

engines in different scenarios based on user data,

consumer behavior data and commodity data, support

users with different roles to call different

recommendation strategies through the Web server of

E-commerce platform, and provide personalized

information services and consumer decision support

for E-commerce platform consumers.

2 OVERVIEW OF KEY

TECHNOLOGIES

2.1 Data Source

The core function of E-commerce platform

recommendation service system is to mine users'

interests and preferences, predict users' implicit needs

and explicit needs, and finally make personalized

recommendations for users according to various

actions triggered and executed by consumers on the

platform. As a prerequisite for the function realization

of the data source recommendation system, its quality

and processing mode directly determine the running

effect of the subsequent algorithm model. In the

process of data source selection, common elements

involve four parts: user information, commodity

information, user consumption behavior and user

scene. The examples covered are shown in Table 1.

Table 1: Common data information table of the E-commerce platform.

No.

Data elements Example

Userinfo Age, gender, education background, occupation, family composition, etc

Commodity information Categories, brand, price, origin, weight, color, specification, unit, shelf life, etc

User consumption

behavior

Browse, find, click, play, add shopping cart, collect, comment, retweet, etc

User's scene Geographic location, system interface, time, specific festivals, major events, etc

The user information and commodity information

come from the internal database of the E-commerce

platform, while the user's consumption behavior and

scene come from the user log file of the E-commerce

platform. In addition, according to the characteristics

of the E-commerce platform itself, the four data

elements include structured data, semi-structured data

and unstructured data. Semi-structured data and

unstructured data can't be directly input into the

recommended algorithm model for operation, so ETL

(Extract-Transform-Load) tool is needed to complete

data preprocessing, and after feature engineering, it

finally meets the operation standard.

Generally, data preprocessing involves three

stages: extraction, conversion and loading, aiming at

integrating scattered, messy and inconsistent data in

E-commerce platform, and transforming unstructured

or semi-structured data into structured data, which

will facilitate the subsequent data application.

Extraction is the process of data collection, that is, the

aggregation of all kinds of data required by the

recommendation service system. For structured data,

you can directly call or build a cross-reference map to

complete the collection, while the log files need to be

buried in the client interface and obtained by the log

collection Web server. In the transformation stage, all

kinds of data information should be cleaned, format

adjusted, missing filled, deleted and repeated, and

finally a data with uniform format, high structure,

high data quality and good compatibility can be

obtained. (Chen, 2016) In the final loading stage, the

converted data can be transmitted to the

recommendation system for storage and provided to

the feature engineering stage of the recommendation

algorithm for processing.

In order to make the recommendation results of

recommendation service system more in line with

users' real needs and satisfy users' personalized and

differentiated consumption psychology, it is

necessary to carry out feature engineering processing

according to various data to form a unique user

portrait model. The central idea of user portrait model

is to extract users' hidden interests and preferences,

and form labels to help the accuracy of personalized

Design and Application of Personalized Recommendation Algorithm Model Based on E-Commerce Platform Data

311

recommendation services. The user portrait model

will integrate four data elements: user information,

commodity information, consumer behavior and user

scene. The modeling results are shown in Figure 1,

where U, A, I and T represent the collection of

information elements, while u, a, i and t represent the

specific information content.

Figure 1: User portrait model structure diagram.

2.2 Recommendation Algorithm

Recommendation algorithm is the core content of

recommendation service system, and it is the key to

realize personalized recommendation technology.

The quality of recommendation algorithm design

fundamentally determines the effect of

recommendation service, and also relates to the

application effect of recommendation service system

in E-commerce platform. In the actual application

process, there are many algorithms that can meet the

personalized recommendation service of E-

commerce platform, and their design ideas, starting

points and data analysis and processing processes are

different. Common algorithm models include

commodity association rule algorithm, demographic

information rule algorithm and collaborative filtering

recommendation algorithm.

Commodity-based association rule algorithm is a

conventional recommendation technology, which can

be widely applied to various E-commerce platforms.

The principle is to generate association rules by

analyzing the internal relations of all commodities

added to the shopping cart in users' consumption

behavior. When a certain commodity appears in the

shopping cart of users, an associated commodity

recommendation list is automatically generated, thus

improving the cross-selling ability of commodities.

The specific association rules are shown in Formula

1, where S represents the goods that trigger the rules,

H represents the recommended goods, P represents

the support, and C represents the confidence.

()

C%p%,HS 

(1)

The algorithm based on demographic information

rules is a relatively simple recommendation

technology, which is suitable for new E-commerce

users who have no historical consumption behavior.

The user groups are classified only by the differences

of user information, and then corresponding

recommendation strategies are formulated according

to different classifications. Most of the user

information is text fields, which belong to

unstructured data. When classifying information, it is

necessary to use TF-IDF, Naive Bayes and other

methods to complete feature extraction before

realizing classification.

Collaborative filtering-based recommendation

algorithm is the most widely used recommendation

technology at present. The theoretical basis of

collaborative filtering is that each user's interest

preferences remain relatively stable within a certain

time range, and each user is not independent, and can

form a small group by virtue of the similarity of

interest preferences. Therefore, the interest

preferences of other members can be predicted by the

interest preferences of one group member. (Liu,

2022) The implementation steps of recommendation

algorithm based on collaborative process include data

acquisition, nearest neighbor search and

recommendation set generation. The user images

formed after data collection and processing can

represent users' interests and preferences. In the

process of nearest neighbor search, two algorithms,

user-based collaborative filtering (User-CF) and

commodity-based collaborative filtering (Item-CF),

can be constructed by using Euclidean Distance,

Pearson Correlation and Cosine Similarity. After the

algorithm is calculated, the similarity is sorted to form

a recommendation list, and the results of TOP-N

items are pushed to users.

Compared with the three recommendation

algorithms, the association rule algorithm based on

commodities is more convenient, but it can't adapt to

the analysis and processing of massive data, and the

degree of personalized recommendation is low, and

the recommendation accuracy is insufficient.

However, the algorithm based on demographic

information rules has a rough overall processing

ICEMME 2022 - The International Conference on Economic Management and Model Engineering

312

process, and the actual effect is not good. The

integrity and authenticity of user information have a

great influence on the recommendation results.

Collaborative filtering recommendation algorithm

has a high degree of automation and intelligence,

which can adapt to large data sets and complete

complex personalized recommendations, but it lacks

good performance in dealing with the cold start of

new users. On the whole, we can combine the rules of

demographic system with collaborative filtering

recommendation algorithm to achieve the purpose of

complementary advantages, and finally improve the

quality of personalized recommendation service.

2.3 Development Process

According to the above application requirements, we

complete the configuration and deployment of the

personalized recommendation service system

development environment. The development content

of the system is divided into two parts, one is the

construction of hardware equipment and

development environment, the other is the

construction and training of each recommendation

algorithm model in Python environment by using data

processing class libraries such as Numpy and Pandas,

forming a personalized recommendation system that

can support Web Server calls.

First of all, in terms of hardware device selection,

according to the system application requirements, the

CPU selects intel core-i7-10700F @ 2.90GHz, 8

cores, 16GB memory and 1TB hard disk space. In the

process of building the software environment, the

operating system is Linux CentOS 4.7, and the

compiling environment is Visual Studio 2019. Pyton

4.2.0 is selected as the development environment, and

Anaconda integrated distribution can be selected for

installation and deployment to shorten the time. In

addition, the deployment and import of Numpy and

Pandas need to be completed with the help of

PyCharm tool, and after verification and

configuration, the configuration of development

environment can be completed.

Secondly, under PyCharm, according to the

implementation process of personalized

recommendation service, we complete data import,

data preprocessing, user portrait model construction

and related algorithm construction in turn. For

example, in the process of constructing collaborative

filtering algorithm, the design and development of

User-CF and Item-CF will be completed according to

user information and commodity information

respectively. The key code for implementing the

User-CF algorithm model is shown in Figure 2, and

the similarity calculation method adopts cosine

similarity evaluation.

Figure 2: Implementation code of user-based collaborative

filtering algorithm (User-CF)

After the recommended algorithm model is built,

a suitable training path is made in the train.py file, and

a large amount of data is imported to complete the

model training. After the whole system is developed,

it will be packaged and published on the Web Server

server of E-commerce platform. After the

corresponding ports are configured, users can use the

system from the client browser. Through the

introduction of the above key technical theories, the

construction environment, related software and

implementation process of personalized

recommendation algorithm are determined, and the

technical feasibility of the whole project of

personalized recommendation service system based

on E-commerce platform is also clarified.

3 DETAILED FUNCTION

IMPLEMENTATION

3.1 Data Collection Module

When the user logs in to the E-commerce platform,

the personalized recommendation system will

automatically start, and obtain the user information,

the user's historical consumption behavior and the

scene information of the user. After data

preprocessing, a user portrait model is formed, and

the corresponding interest preference features are

input into the user feature database to be saved.

Figure 3 shows the key code of missing data

execution processing in the data preprocessing stage.

Design and Application of Personalized Recommendation Algorithm Model Based on E-Commerce Platform Data

313

Figure 3: Missing key code of data processing in data

preprocessing stage.

This process makes a basic distinction between

new customers and regular customers. Because the

new customers lack historical consumption data, the

algorithm called in the follow-up personalized

recommendation is different from the regular

customers, further improving the three-dimensional

and comprehensive personalized recommendation

service.

3.2 Personalized Recommendation

Module

When users realize information retrieval and

browsing through E-commerce platform,

personalized recommendation module will

automatically call various recommendation algorithm

models to complete personalized product information

recommendation. The realization of this function will

also be differentiated according to the roles of new

customers and regular customers. Regular customers

can combine User-CF and Item-CF recommendation

algorithms, and improve the recommendation quality

by weighting, transformation, combination, feature

combination, cascading and other combinations. In

addition, for new customers, personalized

recommendation can be completed by using the rule

algorithm based on demographic information alone,

as shown in Figure 4, which is the key code for

completing user clustering according to K-means

algorithm in Python environment.

Figure 4: K-means clustering algorithm based on

demographic information rules.

3.3 Personalized Interface

The personalized interface will complete the display

of personalized recommendation results, and the final

results will be presented in the form of product list.

At the same time, this list is fed back to customers

through the server, so that customers can know the

personalized recommendation information in real

time, improve their online purchase desire, and

improve the sales performance of E-commerce

platform. (Wang, 2019) After testing, the

collaborative filtering algorithm and demographic

information rule algorithm supported by the system

can well meet the recommendation service function

requirements of the general e-commerce platform. As

shown in Table 2, the system shows the performance

of Item-CF and User-CF algorithms in the test data

set.

Table 2: The performance of Item-CF and User-CF algorithms in the test data set.

Number of neighbors

(K)

Item-CF User-CF

Similarity

(Nsim)

Neibr ratio

(Neibr)

Similarity (Nsim)

Neibr ratio

(Neibr)

10 0.6241 0.3501 0.5607 1.0000

20 0.5884 0.3581 0.5274 0.9988

40 0.5657 0.3461 0.4766 0.9733

4 CONCLUSIONS

In order to realize the personalized recommendation

system of E-commerce platform, the author takes the

personalized recommendation algorithm model as the

core, selects collaborative filtering algorithm as the

main network model, and matches the association

rule algorithm and demographic information rule

algorithm. Under Python development environment,

the author relies on Numpy and Pandas data

processing class library to realize the construction and

training of personalized recommendation service

intelligent machine, and supports the call of E-

ICEMME 2022 - The International Conference on Economic Management and Model Engineering

314

commerce platform Web server. The system starts

automatically when users log in, predicts the products

suitable for users and recommends them according to

user behavior data information, promotes the

effective transformation of users' implicit needs,

enhances the cross-selling ability of E-commerce

system, and provides higher consumption experience

for online shoppers.

ACKNOWLEDGEMENTS

The Project of innovation and entrepreneurship for

college students of Liaoning province

(No.S202213198005).

REFERENCES

Chen Zongyan, Yan Jun. (2016) Collaborative Filtering

Recommendation Algorithm Based on Sparse Data

Preprocessing. Computer Technology and

Development.06:59-64.

Du Huiyuan. (2018) Research on the Impact of Information

Overload of E-commerce Platform on Consumers'

Purchase Decision. University of International

Business and Economics. 05.

Liu Hualing, Guo Yuan, et al. (2022) Research Progress of

Similarity Algorithm in Collaborative Filtering.

Computer Engineering and Applications.04:27-33.

Shang Mengying, Bai Zong, et al. (2021) Design and

Implementation of E-commerce Personalized

Recommendation System. Electronic Technology &

Software Engineering.12:128-129.

Wang Min. (2019) Design and Implementation of E-

commerce Personalized Recommendation System

Based on Collaborative Filtering. Guangzhou

University.05.

Design and Application of Personalized Recommendation Algorithm Model Based on E-Commerce Platform Data

315