RESEARCH ON E-COMMERCE CUSTOMER

SATISFACTION EVALUATION METHOD

BASED ON PSO-LSTM AND TEXT MINING

Qin Yang*

School of Marxism, Jinling Institute of Science and Technology, Nanjing, Jiangsu,

211169, China

yang2161980@sina.com

Reception: 13/11/2022 Acceptance: 29/12/2022 Publication: 13/01/2023

Suggested citation:

Y., Qin (2023). Research on E-commerce Customer Satisfaction Evaluation

Method Based on PSO-LSTM and Text Mining. 3C Empresa. Investigación y

pensamiento crítico, 12(1), 51-66. https://doi.org/10.17993/3cemp.2023.120151.51-66

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

ABSTRACT

With the increase of social technology, e-commerce platforms have entered a period

of rapid development. Improving customer satisfaction and purchase rate is the key to

the survival of e-commerce platforms. Text mining and analysis of customer

evaluation data will help to grasp the focus of customers and optimize the e-

commerce platform. To this end, through text mining technology, the text comment

data of five e-commerce platforms such as Amazon, eBay, Alibaba, Jingdong, and

Taobao are collected, and the cleaned text is analyzed by particle swarm algorithm

(PSO)-long short-term memory (LSTM) model. The data is subject to time scale

extraction, and the extraction results are visualized and interpreted. The research

shows that the logistics, price, freshness, quality and packaging of e-commerce

platform merchants are important factors that affect the evaluation of e-commerce

customer satisfaction.

KEYWORDS

text mining; PSO-LSTM; particle swarm algorithm; long short-term memory network

PAPER INDEX

ABSTRACT

KEYWORDS

1. INTRODUCTION

2. RESEARCH ON TEXT MINING AND ANALYSIS METHODS

3. E-COMMERCE CUSTOMER SATISFACTION EVALUATION

METHOD BASED ON PSO-LSTM AND TEXT MINING

4. CONCLUSION

REFERENCES

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

1. INTRODUCTION

Online shopping has become an important form of shopping. Not only are more

and more consumers choosing to shop online, but also more and more categories of

consumers are shopping online. After several years of development, various e-

commerce platforms are in full swing, but fresh food e-commerce, known as the "last

blue ocean" in the e-commerce industry, is still in its infancy. The biggest difference

between fresh e-commerce and other e-commerce is that the products pay more

attention to freshness and are not easy to preserve. Data from Research shows that

the e-commerce market is developing rapidly, with an average annual growth rate of

more than 50%. However, due to the constraints of many factors, the overall service

of e-commerce is still in the immature stage, its business model is still developing, and

the service level is also mixed. The degree is not optimistic [1-3]. The epidemic in

2020 has brought great opportunities to e-commerce [4-6]. Therefore, how to seize

this opportunity requires in-depth analysis of the factors affecting customer

satisfaction on e-commerce platforms, improving e-commerce products and services,

increasing customer satisfaction, and enhancing user stickiness, thereby promoting

the development of e-commerce.

For the analysis of e-commerce customer satisfaction evaluation methods, scholars

at home and abroad mainly use the text mining method based on sentiment analysis,

classify the vocabulary contained in the customer evaluation content, and quantify the

emotional trend in the text content through weighted assignment. Based on this

method, Susan et al. established an evaluation help degree model to describe the

enthusiasm of the evaluation, but did not consider the difference between the

evaluation of professional buyers and ordinary customers, and lacked accuracy. The

problem of simple evaluation content sentiment analysis has been basically solved,

but the evaluation of evaluation content is affected by the additional functions of the e-

commerce platform, and a variety of restrictive factors must be comprehensively

considered. Yang Ligong et al. [7] used Markov Logic Network to combine sentence

context and emotional features for sentiment analysis, and realized cross-domain text

analysis. Ming Junren [8] applied the association rule method to text mining analysis,

and designed a text data analysis method integrating semantics and association

mining, which improved the accuracy of text analysis. Cai Xiaozhen et al. [9] selected

4 indicators as the basis for online user comments, and constructed a text mining

model to solve the problem of uneven quality of online user comments. Tang Xiaobo

et al. [10] combined co-word analysis and polarity transfer method for text analysis,

which overcomes the shortcomings of traditional text analysis and has better text

analysis results. Liu L[11] and others proposed to combine the feature vector model

and the weighting algorithm of product review sentiment analysis for text analysis.

Walaa Medhat [12] proposed that the conventional analysis steps for commodity text

analysis are commodity review, emotion recognition, feature selection, emotion

classification, and emotion polarity judgment. The above studies have improved the

accuracy and efficiency of text analysis, and laid the foundation for the use of particle

swarm algorithm to mine intelligence value in text.

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

With the rapid development of society, accurate e-commerce customer satisfaction

prediction is becoming more and more important. The accurate prediction of e-

commerce customer satisfaction not only plays an irreplaceable role in the long-term

stable operation of e-commerce platforms and sellers, but also plays an important role

in reducing the cost of e-commerce platforms, improving product quality and market

planning. With the emergence of deep learning, many scholars have turned their

attention to deep belief networks, convolutional neural networks, and recurrent neural

networks. The related research progress is shown in Figure 1.

Figure 1 Research progress of LSTM model

As a special form of regular recurrent neural network (RNN), long short-term

memory network (LSTM) was first proposed by Hochreiter et al. in 1997, and it is

widely used in text mining data prediction. In order to improve the prediction accuracy

of e-commerce customer satisfaction, particle swarm optimization (PSO) optimization

of long short-term memory (LSTM) neural network hyperparameters for e-commerce

customer satisfaction prediction model (PSO-LSTM) is widely used. Aiming at the

problem that LSTM hyperparameters are difficult to select, the PSO algorithm can

effectively find the global optimal solution to optimize the hyperparameters of the

LSTM model, and continuously train to find suitable hyperparameters and verify them.

Therefore, this paper combines the PSO-LSTM algorithm with text mining technology

to study and analyze the factors that affect the evaluation of e-commerce customer

satisfaction.

2. RESEARCH ON TEXT MINING AND ANALYSIS

METHODS

Online reviews of e-commerce customers are messy but valuable unstructured or

semi-structured data. The information contained in online reviews has an important

impact on the strategic decision-making of merchants and the purchasing decisions of

consumers [13-14]. At present, the academic community has recognized the

importance of online reviews, but the research on online review willingness has not

received much attention. It is necessary to link online reviews with Internet word-of-

mouth, and study consumers' review willingness from the perspective of Internet

word-of-mouth dissemination willingness. This paper directly studies the influencing

factors of customer satisfaction evaluation, and uses the method of text mining in the

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

research to explore which factors will affect consumers' evaluation behavior and

willingness [15].

Text mining analysis has a strong foundation under the support of network data.

Sem Eval, an international evaluation expert, defines evaluation objects as

expressions that can be used to express the characteristics of evaluation entities in

specified texts. When evaluating an entity as a commodity, the evaluation object at

this time may be the product features, functions, and parts corresponding to the entity.

For product reviews on e-commerce platforms, the evaluation objects include express

delivery, services, etc. in addition to the inherent attributes of the product. For

example: "The phone looks tall and the system runs very fast, but the power is not

durable, so I'm a little disappointed." For this comment, "appearance" and "battery"

are the evaluation objects, "tall", "Running very fast" and "Not durable" are the

comments corresponding to the three evaluation objects. Therefore, the extraction of

evaluation objects and comments plays an important role in the analysis of commodity

evaluation texts. By analyzing the above comments, it can be found that the

comments corresponding to the first two evaluation objects are positive, while the

comments corresponding to the third evaluation object are negative. In the actual

analysis process, whether the reviews are classified as positive, negative or neutral

will affect the accuracy of text mining analysis of product reviews [16-17]. Therefore,

the text mining process will extract the evaluation objects and their corresponding

comments in each product review, and analyze them, which increases the accuracy of

text mining analysis to a certain extent, and can be better applied to e-commerce

customer satisfaction. Under evaluation. In addition, it helps e-commerce companies

understand their competitive environment and position, discover the breakthrough

points of the industry, and further adjust their own development strategies, so that

consumers' purchasing experience in merchants can be optimized, and it is also

helpful to improve the overall industry. performance level. Taking computer furniture,

seafood and aquatic products, fruits and vegetables, and daily necessities as

examples, some data of specific customer full score evaluation are shown in Table 1.

Table 1 Some product review data

E-commerce customer Comment product category

M***P

I received the phone, and the

pink color is too beautiful. I

really like this phone. Can

last all day on a single

charge

phone

Little***j

The logistics is very fast, and

the lining protection of the

packaging is very charac-

teristic. The dual graphics

cards work together, and the

video and audio effects are

good~~~

computer

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

The process of text mining analysis method includes text feature extraction, text

data cleaning, high-frequency text selection and processing. The feature

representation of text refers to a method that can use words, words, phrases or

sentences as feature items to represent the entire text, and complete unstructured

text processing by processing these feature items. Text features can generally be

accurately expressed by words, words and phrases, while sentences and paragraphs

can be further divided into words, words and phrases. The text representation method

is currently uncommon. Commonly used text feature representation types are as

follows:

(1) Words. Individual numbers, letters, spaces, special symbols, and Chinese

characters that constitute the structural units of phrases, phrases, and concepts

cannot reflect the characteristic meaning and emotional expression of the text, and

need to be effectively combined. Therefore, word-based text feature representation

requires Extraction and optimization of text words [18-19].

(2) Phrases. To a certain extent, it can be used as a combination of words at the

most basic semantic level. Phrase-based text feature representation In a certain

limited field, the phrase feature space may contain tens of thousands or even millions

of phrases, which requires special dimensionality reduction processing.

Call***girl

The vegetables are very

fresh and the price is affor-

dable. There are also garlic

moss and eggplants that are

currently scarce in the ma-

rket. It is really good. Thank

you for the efforts made by

the ** platform for us!

vegetable and fruit

OH***！

The platform is fast, I place

an order in the morning and

arrive in the afternoon. After

the logistics arrives, the

frozen state is kept well, and

the fish is very fresh.

seafood

See***door

My son loves reading books

very much, he bought a lot

of books on the e-commerce

platform, the quality is very

good, very beautiful

book

MY***baby

All solid wood furniture, no

glue, no paint, with a touch

of pine fragrance, I like it

very much

furniture

YOU***left

The fabric and form are

really, really good! Bought it

for my husband, the e-

commerce customer service

is very greasy and loving!

Buy with confidence!

clothing

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(3) Phrases. Single words or compound phrases extracted from the original corpus

directly by entity extraction methods are generally composed of specific words, and

most of them exist in typical text dictionaries.

(4) Concept. Based on rules or hybrid classification methods, through

preprocessing procedures, specific unit combinations are formed by manual counting,

identifying individual words, compound phrases, entire sentences and even larger

syntactic units.

This text mining analysis collects text comment data from five e-commerce

platforms such as Amazon, eBay, Alibaba, JD.com, and Taobao. The product types

include furniture, home appliances, digital products, fresh fruits, seafood and meat,

and daily necessities[20-22]. The collected text data is processed by data cleaning,

word segmentation, etc., and text mining analysis is performed on the data to

construct high-frequency words in comments to study the customer satisfaction

evaluation of e-commerce platforms.

1) Use web crawler technology to capture customer comment texts of merchants

on various e-commerce platforms, filter and clean the data, remove word

segmentation and other preprocessing.

2) Perform text mining analysis on the preprocessed data. This stage mainly

includes the following steps: ①

Select an appropriate method to determine the

number of topics (because the number of topics is too large or too small will affect the

interpretation of text analysis results); ②

Use text mining analysis to extract topics,

and thus generate document-topic Distribution matrix and topic-vocabulary distribution

matrix; ③Simple sorting and interpretation of the extracted topics; ④

Visually display

the results of text analysis.

3) Through the interpretation of the text analysis results, it is finally verified how the

merchants of the e-commerce platform affect the satisfaction evaluation of

consumers, and whether there are other factors in the influence process that mediate

or moderate the influence process [23-24].

In order to study the e-commerce customer satisfaction evaluation criteria more

accurately, the text topics identified by text mining analysis are in common: logistics,

price, service, quality, and packaging. Based on the five types of e-commerce

platforms such as Amazon, eBay, Alibaba, Jingdong and Taobao, the text analysis

method is used to calculate the five text themes in the five categories of furniture,

home appliances, digital products, fresh fruits, seafood and meat, and daily

necessities. The weight occupied, and the radar chart is drawn according to the

weight, as shown in Figure 2.

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Figure 2 Radar chart of various themes of various e-commerce platforms

It can be seen from Figure 2 that customers value different characteristics of

different types of products on the e-commerce platform. For example, for commodities

such as fresh fruits, seafood and meat, consumers place the most importance on

freshness and quality, and for other types of products, consumers place the most

importance on quality. On the whole, consumers pay more attention to the quality,

price and logistics of the products of the e-commerce platform. According to the

research results, for the major e-commerce platforms, the following suggestions are

put forward to improve customer satisfaction:

quality. The above e-commerce platforms all have stable brand suppliers and

controllable product sources. There are certain advantages in terms of quality. E-

commerce platforms can also achieve quality traceability through technologies such

as blockchain and the Internet of Things. After the information is visualized,

consumers have more trust in the quality of the products, and the platform will pay

more attention to the quality problems in the production process. Price [25]. Some

products of the above-mentioned e-commerce platforms are directly sourced, and

some products have established their own product bases, so they can provide

relatively favorable prices. E-commerce platforms can also compare multiple parties

when selecting suppliers. Under the premise of ensuring quality, they can cooperate

with suppliers for a long time. At the same time, they can optimize logistics,

warehousing and other solutions to reduce costs, so as to give customers the best

price.

3. E-COMMERCE CUSTOMER SATISFACTION

EVALUATION METHOD BASED ON PSO-LSTM AND

TEXT MINING

In the previous section, we conducted text mining on e-commerce customer

satisfaction evaluation texts to explore how e-commerce platforms affect consumer

satisfaction evaluation behavior, and to find important factors that affect e-commerce

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

customers’ satisfaction evaluation behavior. The factor is single, only high-frequency

words are counted, and the time changes of customer evaluations are not considered,

so the accuracy is low. The PSO-based LSTM algorithm transforms its structure on

the basis of the traditional artificial neural network algorithm, so as to achieve the

purpose of enabling the network to remember past time information, so that the

network can not only realize the connection from bottom to top (input-output), but also

It can also realize information transmission and recording from left to right (time t-time

t+1). Therefore, the combination of PSO-LSTM calculation and text mining analysis

can more accurately study the satisfaction evaluation of e-commerce customers. The

algorithm flow of PSO-LSTM will be discussed in depth below.

1 The forward propagation process

(1) Input-hidden layer

The PSO-LSTM algorithm is a two-dimensional feature algorithm. After the input

layer-hidden layer-output layer operation, the local features of the input data are

extracted. In this paper, ft is used to represent the channel input of the PSO-LSTM

algorithm model, yt is the output weight of the PSO-LSTM algorithm model, Wf is the

bias vector of the input layer, and Vht is the receptive field of the output layer on the

input data. The algorithm process is shown in formulas (1)-(2):

(1)

(2) Hidden layer - output layer

(2)

2. Backpropagation process

(1) Define the error function of the sequence

At time , the actual output of the LSTM is , while the

expected output is ，Then the loss function of the

entire time series is

(3)

Therefore，

(4)

(2) Define the error term at time t

That is, the partial derivative of the loss function to the output value. First, define

the error term of the output layer at time t as ，Second, define the error

term in the hidden layer at time t.

( )

t f t f t f

f W h U x bσ−

= + +

( ).

t t y

y Vh bσ= +

yt=

{

y1(t),y2(t),…,yp(t)

{

d1(t),d2(t),…,dp(t)

L L

=∑

( )

t t t

L y d=−

yt=

ժL

ժy

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Error in output gate：

(5)

Because the information of all moments is stored in the memory cell state, the error

term 

needs to be accumulated at each moment, and the error returned to the

memory cell is divided into two parts, the first part is returned from the output ℎ of the

hidden layer The second part is the error returned from the memory cell state at the

next moment +1：

(6)

First calculate the error term of the memory cell state in the hidden layer at the

previous moment：

(7)

From this formula, the error returned from the memory cell state at time t+1 +1:

(8)

Combining (6) and (8), the error term of the memory cell state is obtained as:

(9)

Error in input gate:

(10)

Error in output gate:

(11)

(3) Calculate the error gradient of the weight coefficient matrix

The first step is to calculate the error gradient of the weight matrix 

from the

hidden layer to the output layer, and the two sides of the loss function 

take the

partial derivative of  to get:

(12)

( )

1 ,

tanh .

t t t

t t t t

t t t

L L y

h y y y V

h y h

L L h

o h C

o h o

δ δ

∂ ∂ ∂

= = = −

∂ ∂ ∂

= = =

∂ ∂ ∂





( )2

a .1 t nh

t t t

t t t t

t t t

L L h

C h C o

C h C

δ δ

∂ ∂ ∂ ⎡ ⎤

= = = −

⎣ ⎦

∂ ∂ ∂  

1 1

t t t

L L C

C C f

C C C

δ δ

−

− −

∂ ∂ ∂

= = =

∂ ∂ ∂ 

1 1

t t t

C C fδ δ

+ +

=

( )

1 1

t .1 anh

t t t t t t

C C f h C oδ δ δ

+ +

⎡ ⎤

= + −

⎣ ⎦



t t t

t t

L L C

i C C

i C i

L L C

C C i

C C

δ δ

∂ ∂ ∂

= = =

∂ ∂ ∂

= = =

∂

∂ ∂





 

t t t

L L C

f C C

f C f

δ δ −

∂ ∂ ∂

= = =

∂ ∂ ∂ 

( )

1 1 1

T T T

t t t

t t t t

t t t

L L y

Ly y y h

V V y V δ

= = =

∂ ∂ ∂

∂= = = −

∂ ∂ ∂ ∂

∑ ∑ ∑  

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

The next step is to calculate the error coefficient in the hidden layer:

(13)

The error gradient from the input layer to the output layer weight matrix can also be

obtained in the same way:

(14)

(4) Update of the weight matrix

(15)

In this paper, the PSO-LSTM algorithm uses an activation function in the neurons

and the final prediction layer, which can combine the linear input nonlinearly, so that

the PSO-LSTM algorithm has nonlinear factors, and the prediction is more accurate.

The commonly used activation functions are Sigmoid function, Relu function, Tanh

function. The activation functions of the LSTM network at the neurons are the Relu

function and the Tanh function, which are often used because of their simple

calculation and fast iteration speed [26-27]. The Sigmoid function is used in the output

layer for binary classification probability calculation.

The mathematical formula for the sigmoid function is：

(16)

The mathematical formula for the Tanh function is:

(17)

The mathematical formula for the Relu function is：

(18)

According to the traditional method of training deep models, the LSTM model in this

paper uses the stochastic gradient descent algorithm to train parameters. Choose a

cross-entropy loss function according to the classification target definition：

(19)

Among them, N represents the number of samples, and c represents the number of

categories. tkn represents the true category of the nth sample, and ykn represents the

prediction result of the nth sample. The purpose of training the model is to minimize

the loss function. Define the output of layer Z as：

( )

1 1 1

1 .

T T T

t t t

t t t t

t t t

f f t f

L L f

Lf f f h

W W f W δ

−

= = =

∂ ∂ ∂

∂= = = −

∂ ∂ ∂ ∂

∑ ∑ ∑

 

( )

1 1 1

1 .

T T T

t t t

t t t t

t t t

f f t f

L L f

Lf f f x

U U f U δ

= = =

∂ ∂ ∂

∂= = = −

∂ ∂ ∂ ∂

∑ ∑ ∑  

new

V V

η∂

= + ∂

( ) 1x

sigmoid x e−

tanh( ) 1

−

( ) max(0, )f x x=

1 1

1log

N c

N n n

k k

n k

E t y

N= =

=−∑∑

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(20)

Among them, f represents the activation function, xl−

1 represents the output of the

L-1 layer, and for the L layer, it is also its input, l represents the weight of the L layer,

and b represents the L layer bias. In the process of error back propagation algorithm,

this model uses gradient descent method to update the weight of the calculation

network. The update calculation formula of gradient descent method is as follows:

(21)

Among them, represents the learning rate in the gradient descent calculation, and

the update formula of the weight of each layer can be obtained according to the chain

derivation rule. In the process of classifying information texts, there are many

situations in the classification results. The following four possibilities exist for the

classification results:

(1) True class True positives (TP): The number of positive samples that were

successfully identified;

(2) False positives (FP): The number of negative samples that are misidentified;

(3) False negatives (FN): The number of positive samples that are incorrectly

identified;

(4) True negatives (TN): The number of negative samples that are correctly

identified;

Based on the method of text mining in Chapter 3, and the high-frequency

vocabulary logistics, price, freshness, quality and packaging it excavated are the text

positive classification data of PSO-LSTM. The sample data are customer evaluations

of furniture appliances, digital products, fresh fruits, seafood and daily necessities of

the five major e-commerce platforms, from March 2017 to March 2021. The essence

of the experiment in this paper is to classify information texts, so only the text and

category labels in the collected forum data are extracted as experimental data. The

purpose of selecting category labels is to distinguish normal information and spam

information with 0 and 1 respectively, the class labels in the category are verified

manually, so the labels of the data are true and reliable. The total number of collected

data is 83,961 texts, including 14,678 spam messages and 69,283 normal messages.

Formulas (1)-(15) are the calculation equations for filtering text data by PSO-LSTM,

and the activation function is used in the neurons and the final prediction layer, which

can combine the linear input nonlinearly, so that the PSO-LSTM algorithm has non-

linear combination. Linear factor, the prediction is more accurate, and the high-

frequency words of text mining are classified and calculated according to formulas

(22)-(25). Through the previous theoretical research basis and data processing

( )

l l

l l l l

x f u

u x bω

−

= +

new old

old

new old

old

l l

b b

ω ω η ω

∂

=−∂

∂

=−∂

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

process, the time division of the weights of high-frequency vocabulary logistics, price,

freshness, quality and packaging is carried out, as shown in Figure 3

Figure 3 The relationship between the characteristics of various e-commerce

platforms (a) furniture and home appliances (b) digital products (c) fresh fruit (d)

seafood and meat (e) daily necessities over time

As can be seen from Figure 3(a), with the increase of time, the highest change rate

of high-frequency words in e-commerce customers' satisfaction with furniture and

home appliances is logistics speed, and its weight is 52 in 2017, and gradually

increases to 91 in 2021. This also means that when other factors remain unchanged,

the faster the logistics of furniture and home appliances on the e-commerce platform,

the higher the customer satisfaction rating; as shown in Figure 3(b), as time goes by,

e-commerce customers are more interested in The highest rate of change in the high-

frequency words of satisfaction of digital products is also the logistics speed, and its

weights from 2017 to 2021 are 51, 54, 73, 89 and 95 respectively; as can be seen

from Figure 3(c)-(d), With the increase of time, the most frequent change rate of e-

commerce customers' satisfaction with fresh fruits, seafood and meat is their

freshness, followed by logistics. The weight of the freshness of the subject word was

82 in 2017 and gradually increased to 97 in 2021. This also means that when other

factors remain unchanged, the faster the freshness of fresh fruits and seafood on the

e-commerce platform, the higher the customer satisfaction rating; as shown in Figure

3(d), with the increase of time, E-commerce customers' satisfaction with daily

necessities has the highest change rate of high-frequency words in price, followed by

quality.

In summary, in 2017, the development of e-commerce platforms was relatively

weak, the number of e-commerce shoppers was relatively small, and customer

satisfaction at that time was more inclined to price, which also meant that the higher

the product prices of the merchants on the e-commerce platform. When it is low, the

customer satisfaction rating is higher. With the development of e-commerce platforms,

customers pay more and more attention to the speed of logistics and the freshness of

products. Through data processing through the PSO-LSTM algorithm, it can be seen

that in 2018, the proportion of logistics and freshness has increased significantly,

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

which also means that e-commerce The faster the product logistics speed of the

merchants on the platform and the higher the freshness, the higher the customer

satisfaction rating. With the advent of the epidemic, the e-commerce platform has

entered an ice age, but through data analysis, it can be seen that from 2019 to 2021,

product orders on the e-commerce platform will still grow steadily.

4. CONCLUSION

This paper mainly studies the analysis process of e-commerce customer

satisfaction evaluation based on PSO-LSTM and text mining, and conducts analysis

and research according to the obtained results. The text data of the text all come from

five categories of products such as furniture, home appliances, digital products, fresh

fruits, seafood and meat, and daily necessities on the five e-commerce platforms of

Amazon, eBay, Alibaba, Jingdong, and Taobao. When using text mining technology to

analyze the evaluation of e-commerce customer satisfaction, five subject words that

have the highest impact on customer satisfaction evaluation are obtained: logistics,

price, freshness, quality and packaging.

Since the text mining technology does not consider the time factor, the accuracy of

the measured factors is low. In order to more accurately study the factors affecting the

evaluation of e-commerce customer satisfaction on the time scale, the text analysis

technology and the PSO-LSTM model were combined to further analyze the five types

of high-frequency words that affect the evaluation of e-commerce customer

satisfaction, and to analyze Its proportion is analyzed. Analyzed the proportion of

factors affecting the evaluation of e-commerce customer satisfaction in each time

period. For furniture and home appliances, with the increase of time, the highest

change rate of high-frequency words in e-commerce customers' satisfaction with

furniture and home appliances is logistics speed, the maximum weight is 91 in 2021,

and the minimum weight is 52 in 2017; E-commerce customers who buy fresh fruit or

seafood are more concerned about the freshness of the goods, and their weight will

gradually increase from 82 in 2017 to 97 in 2021. Through the data processing of

PSO-LSTM algorithm, it can be seen that due to the impact of the epidemic, the

logistics speed and quality have shown a trend of substantial improvement. On the

whole, consumers pay more attention to the quality, price and logistics of the products

of the e-commerce platform. The evaluation method of e-commerce customer

satisfaction based on PSO-LSTM and text mining designed in this paper can

effectively solve the imbalance of information datasets, and is exp

REFERENCES

(1) Le S , Wu Y , Guo Y , et al. Game Theoretic Approach for a service function

chain routing in NFV with coupled constraints[J].

Circuits and Systems II:

Express Briefs, IEEE Transactions on, 2021, PP(99):1-1.

(2) Mitsuru, Toyoda, Yuhu. Mayer-Type Optimal Control of Probabilistic Boolean

Control Network With Uncertain Selection Probabilities.[J]. IEEE

transactions on cybernetics, 2019.

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(3) Wu Y, Guo Y, Toyoda M. Policy Iteration Approach to the Infinite Horizon

Average Optimal Control of Probabilistic Boolean Networks[J]. IEEE

Transactions on Neural Networks and Learning Systems, PP(99):1-15.

(4) Zhang Y, Qian T, Tang W. Buildings-to-distribution-network integration

considering power transformer loading capability and distribution network

reconfiguration[J]. Energy, 2022, 244.

(5) Li, H., Deng, J., Feng, P., Pu, C., Arachchige, D., and Cheng, Q., (2021) Short-

Term Nacelle Orientation Forecasting Using Bilinear Transformation and

ICEEMDAN Framework. Front. Energy Res. 9, 780928.

(6) Li, H., Deng, J., Yuan, S., Feng, P., and Arachchige, D., (2021) Monitoring and

Identifying Wind Turbine Generator Bearing Faults using Deep Belief

Network and EWMA Control Charts. Front. Energy Res. 9, 799039.

(7) Yang Ligong, Tang Shiping, Zhu Jian. Sentence Love Based on Markov Logic

Network Sense Analysis Method [J]. Journal of Beijing University of

Technology, 2013, 33(6): 600-604.

(8) Ming Junren. Research on Text Sentiment Analysis Algorithm Integrating

Semantic Association Mining [J]. Library and Information Work, 2012, 56(15):

99-103.

(9) Cai Xiaozhen, Xu Jian, Wu Sizhu. Research on User Comment Filtering

Model for Sentiment Analysis [J]. Modern Library and Information Technology,

2014, 30(4): 58-64.

(10) Tang Xiaobo, Xiao Lu. Research on Comment Mining Model Based on

Sentiment Analysis [J]. Intelligence Theory and Practice, 2013, 36(7): 100-105.

(11) Liu L, Song W, Wang H, et al. A novel feature-based method for sentiment

analysis of Chinese product reviews[J]. China Communications (English

version), 2014, 11(3): 154-164.

(12) Medhat W, Hassan A, Korashy H. Sentiment analysis algorithms and

applications: A survey[J]. Ain Shams Engineering Journal, 2014, 5(4):

1093-1113.

(13) Hochreiter S, Schmidhuber J. Long Short-Term Memory[J]. Neural Computa-

tion, 1997, 9(8): 1735-1780.

(14) Zhou Ying , Liu Yue , Cai Jun . Sentiment Analysis of Weibo Based on

Attention Mechanism [J]. Intelligence Theory and Practice, 2018,41(03):89-94

(15) Fan Qingbo, Jiang Fucai, Ma Quandang, et al. BP Neural Network-Markov

Ship Traffic Prediction Model Based on PSO [J]. Journal of Shanghai

Maritime University, 2018, 39(2): 22-27, 54.

(16) Feng Yuxu, Li Yumei. Research on CSI 300 Index Prediction Model Based on

LSTM Neural Network [J]. Practice and Understanding of Mathematics, 2019,

49(7): 308-315.

(17) Zhao Jingpeng. Research on hierarchical Web service classification method

based on deep learning [D]. North China Electric Power University (Beijing),

2021.

(18) Luo Zhengjun, Ke Mingsong, Zhou Dequn. Research on Text Sentiment

Analysis Model Based on Improved LSTM [J]. Computer Technology and

Development, 2020, 30(12): 40-44.

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(19) Li Qichao.

Analysis of web data mining technology based on cloud

computing environment [J]. Information Technology and Informatization,

2020(09):54-56.

(20) Kong Jia, Deng Sanhong, Zhang Yue, Zhang Yiwei.

Thematic mining of

science and technology news in the Web environment: Taking the field of

international polar expeditions as an example [J]. Science and Technology

Management Research, 2020, 40(17): 173-179.

(21) Liu Yong.

Design and implementation of substation equipment defect

management system based on Web technology [D]. University of Electronic

Science and Technology of China, 2020. https://doi.org/10.27005/

d.cnki.gdzku.2020.004586.

(22) Yang Mengda.

Temperature prediction based on improved PSO-LSTM

neural network [J]. Modern Information Technology, 2020, 4(04): 110-112.

(23) Xia Maosen, Jiang Lingling. Analysis of "Maker" Hot Words Based on Web

Text Mining [J]. Journal of Hubei University of Science and Technology, 2019,

39(05): 56-62.

(24) He Yun. Research on intelligent traffic flow prediction based on PSO-LSTM

algorithm [J]. Qinghai Transportation Science and Technology, 2019(04):76-79.

(25) Zhou Zhou. Analysis and Research on the Curriculum Setting of Secondary

Vocational E-commerce Majors Based on Web Text Mining [D]. Guangdong

Normal University of Technology, 2019.

(26) Frayssinet, M., Esenarro, D., Juárez, F. F., y Díaz, M. (2021). Methodology

based on the NIST cybersecurity framework as a proposal for

cybersecurity management in government organizations.

3C TIC.

Cuadernos de desarrollo aplicados a las TIC, 10(2), 123-141. https://doi.org/

10.17993/3ctic.2021.102.123-141

(27) Che Xiangbei,Li Man,Zhang Xu,Alassafi Madini O. & Zhang Hongbin.(2021).

Communication architecture of power monitoring system based on

incidence matrix model. Applied Mathematics and Nonlinear Sciences (1).

https://doi.org/10.2478/AMNS.2021.1.00098.

https://doi.org/10.17993/3cemp.2023.120151.51-66

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023