The Construction of a Model for Predicting Users' Repeat Purchase

Behavior and its Impact on the Economic Efficiency of Enterprises

QIAN LYU

Lyceum of the Philippines University Manila Campus,

Manila 1002,

PHILIPPINES

*Corresponding Author

Machine (SVM). Then, the Sparrow Search Algorithm (SSA), based on multi-strategy optimization, is suggested

to overcome the SSVM's drawbacks. The Smooth Support Vector Machine (SSVM) is employed as the feature

classifier for classification. On this basis, an intelligent prediction model of user repeat purchase behavior based

on ISA-SSVM is constructed to achieve efficient prediction of user repeat purchase behavior. The results showed

that the fitness value of the ISA-SSVM algorithm was always higher than other algorithms as the number of

iterations increases. And its convergence speed is fast, when the number of iterations is 13, the fitness value

reaches 94.6%. The error value of this model is 0.14, the loss value is 0.20, the F1 value is 0.957, the recall value

is 0.965, the MAE value is 8.52, the fit degree is 0.992, the prediction accuracy is 97.92%, and the AUC value is

0.995, all of which are better than the other two models. As a result, the ISA-SSVM developed in this work

outperforms previous models in terms of its ability to forecast customers' recurrent purchasing behavior. The

Received: August 15, 2022. Revised: July 14, 2023. Accepted: August 16, 2023. Published: September 21, 2023.

E-commerce has accelerated with the development of

information and Internet technologies and is now one

of the most popular ways to shop. Data relating to user

behavior have been saved in e-commerce platforms in

recent years as the number of businesses using them

grows and the e-commerce industry matures. By

To address this problem, the study proposes to use the

Relief algorithm for feature selection and Smooth

Support Vector Machine (SSVM) as a feature

classifier to classify features. In addition, an improved

Sparrow Search Algorithm (SSA) is used for SSVM

optimization for its unsatisfactory performance, and

finally, an ISSA-SSVM-based intelligent prediction

model of users' repetitive purchase behavior is

constructed to achieve high-precision prediction of

users' behavior and thus improve the economic

optimize SSVM for model accuracy, [1], [2], [3]. The

main structure of the study has four parts: the first part

is to review and analyze the recent related research;

the second part is to construct an intelligent prediction

model of user repetitive purchase behavior using

ISSA-SSVM; the third part is for the model

performance analysis; and the last part is to

summarize the research content.

development opportunities to many e-commerce

enterprises. The prediction of users' repetitive

purchase behavior can help e-commerce enterprises

accurately control users' preferences and purchasing

habits to realize precise marketing and improve

e-commerce enterprises’ benefits, [4], [5]. Therefore,

user repetitive purchase behavior prediction has

received attention from many researchers. To

determine the variables impacting consumer

purchase decisions in omnichannel retailing and to

forecast customer recurrent buying behavior, Mishra

research in China and research trends of SVM and

analyzed the effects of different SVM improvement

methods, providing theoretical support for the

improvement and application of SVM, [15]. Fan and

Sharma combined SVM with least squares support

vector machine (LSSVM) to design an engineering

cost prediction model, to achieve efficient and

accurate intelligent prediction of engineering costs

and provide data support for cost control of

engineering projects, [16]. Neelakandan and Paulraj

[17]. Hamdan designed a statistical SVM-based

intelligent recognition model by which an efficient

and intelligent recognition of handwritten characters

can be achieved and the efficiency of character

recognition can be improved, [18]. [19], used an

improved SVM as a classifier to recognize

dermoscopic images and used the recognition results

to assist in the diagnosis of melanoma. The approach

has a good recognition accuracy and can help

clinically diagnose melanoma, according to the

the SVM model performance and used the improved

significance to improving enterprises' economic

benefits. According to the findings of the most recent

research, repeat purchase intention has been the

subject of numerous studies, although it cannot be

quantified. Moreover, because this kind of intention

dimensionality reduction, [22], [23]. The principle of

KPCA is that based on PCA, a kernel function is

introduced to make the dimensionality of the data to

be processed higher. The effect of KPCA on data

classification is shown in Figure 1.

Calculate the dimensionalized sample covariance

Equation (3).

Fig. 2: The principle of SMOTE algorithm

In Figure 2, the distance between the samples of

a lesser number of species and other samples is first

After obtaining the distance between the

samples of a lesser number of species and other

samples, the imbalance ratio of samples in the

experimental dataset is calculated. Then the sampling

ratio is set by the calculation result

. After

repeated sampling operations by SMOTE algorithm,

it makes all the sampled data in the dataset achieve a

balanced ratio of 1:1. For any few samples

,

within a certain range around it,

samples are

selected to construct a new sample, and the above

Equation (9),

is a random function.

replaces the SMOTE neighbor of the sample in

SMOTE with the center of the few-class sample. It is

hoped that the expanded small class sample will tend

toward the center of the small class sample. Figure 3

shows the data preprocessing process based on the

KPCA-SMOTE algorithm.

In Figure 3, the KPCA-SMOTE algorithm first

3.2 Construction of SSVM based on

Multi-strategy Improved SSA Optimization

To profile user behavior, feature mining is performed

in three dimensions: user, enterprise, and

user-enterprise, and the Relief algorithm can select

features to obtain feature attributes that are more

correlated with user repeat purchase, including user

purchase frequency, purchase conversion rate, and

repeat purchase frequency in that enterprise. These

features are input into the SVM for classification and

identification, and the user's repeat purchase

Fig. 4: The Basic Structure of Support Vector

Machines

The basic principle of SVM for classification of

features is to locate a hyperplane within the feature

space such that the distance between the nearest

points on either side of this hyperplane is maximized,

[24]. To further improve SVM’s classification

accuracy and efficiency, this study improves the

(10)

In Equation (10),

is the weight;

To increase the classification accuracy of the

based on multiple experiments to select the

parameters that make the model performance optimal.

However, this method takes a lot of time and the

selection of the best parameters is often not

satisfactory. To address this problem, a multi-strategy

collaborative improvement SSA is proposed to

optimize the SSVM. The SSA swarm intelligence

optimization technique is used to determine the best

parameter values for the SSVM because it has the

the study proposes a Chebyshev chaotic mapping to

initialize the SSA population, as shown in Equation

(12).

sparrow individual

Chebyshev chaos mapping strategy was able to

increase the diversity of the initial population of SSA,

discoverer individuals, which may lead the algorithm

Based on the above algorithm, the intelligent

repeated purchasing behavior's input and output are

first established. The characteristic amount of the

retrieved data serves as both the model's input signal

and the sort of behavior it will produce. For

normalization processing, the data is separated into

training samples and test samples. The sparrow

was sorted and the individuals with higher fitness

values were the discoverers and the others were the

entrants. In a safe state, the sparrows can search

extensively, and if the population is greater than the

warning value, it will have anti-predator behavior.

Next, update the location of the participant. If the

individual fitness value is relatively low, these users

need to search other locations to improve their fitness.

The global optimal information is updated after

comparing this iteration's fitness value to the best

Fig. 8: Changes in error and loss values

optimizations of SVM parameters were compared to

those of the ISSA optimization. Figure 9 shows the

optimization iteration curve of the cross-validation

accuracy of each algorithm.

Fig. 9: Optimization curve of classification accuracy

Figure 9 illustrates how the fitness of the GWO

algorithm declines after more iterations than that of

other algorithms. When there had been 20 iterations,

the fitness had not altered (82.3%). It demonstrates

that, during the optimization process, the GWO

method is most likely to reach the local optimal. The

the addition of reconnaissance and early warning

mechanisms and dynamic reverse learning

mechanisms, the algorithm has a high classification

accuracy for training samples. Using the ISA-SSVM

of repeat customer behavior is possible. Two metrics,

F1 and Recall, are used to analyze the performance

Figure 11 displays the outcomes of the MAE