RESEARCH ON THE DEVELOPMENT AND

APPLICATION OF CNN MODEL IN MOBILE

PAYMENT TERMINAL AND BLOCKCHAIN

ECONOMY

Hongyan Liu*

Suqian University, School of Foreign Studies, Suqian, Jiangsu, 223800, China

yanyanhong20221123@163.com

Reception: 17/11/2022 Acceptance: 17/01/2023 Publication: 15/02/2023

Suggested citation:

L., Hongyan (2023). Research on the development and application of CNN

model in mobile payment terminal and blockchain economy. 3C Empresa.

Investigación y pensamiento crítico, 12(1), 207-224. https://doi.org/

10.17993/3cemp.2023.120151.207-224

https://doi.org/10.17993/3cemp.2023.120151.207-224

207

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

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

ABSTRACT

As the most widely used payment method at this stage, mobile payment is more and

more closely related to the blockchain economy. Traditional methods lack a certain

degree of accuracy. This research proposes a feature-based and sequential-based

Bilateral AM (BAM) and Convolutional Neural Network (CNN)-gated recurrent unit for

the development and application of mobile payment and blockchain economy (Gated

Recurrent Unit, GRU) hybrid model (BAM-CNN-GRU), select 5 feature parameters

with high correlation with the blockchain for multivariate prediction. The introduction of

BAM can automatically quantify the correlation between the input variables and the

blockchain, and strengthen the expression of historical key information on the

predicted output; the introduction of CNN can extract high-dimensional features that

reflect the non-stationary dynamic changes of the blockchain. The proposed hybrid

model achieves good results in both single-step and multi-step long-term series and

multivariate input blockchain prediction. Compared with the other six methods, MAE is

reduced by 75.45%, 64.74%, 62.84%, respectively. 59.41%, 45.54%, 44.16%.

Compared with the BAM-GRU model, the CNN-GRU model, the GRU model, the

LSTM model, the support vector machine SVM model and the BP model, the

prediction accuracy of the hybrid model has been greatly improved, and it has a

broader application prospect.

KEYWORDS

BAM-CNN-GRU; Mobile payment; Blockchain; Model comparison

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

208

PAPER INDEX

ABSTRACT

KEYWORDS

1. INTRODUCTION

2. PRINCIPLES OF DEEP LEARNING MODELS

2.1. 1D-CNN

2.2. Attention mechanism (AM)

2.3. GRU

3. MOBILE PAYMENT AND BLOCKCHAIN PREDICTION MODEL

BASED ON BAM AND CNN-GRU HYBRID MODEL

3.1. Mobile Payment and Blockchain Prediction Model

3.2. Feature AM

3.3. CNN layer

3.4. temporal AM

3.5. Hybrid model based on BAM and CNN-GRU

3.6. Error Analysis

4. MODEL PREDICTIONS

5. CONCLUSION

6. CONFLICT OF INTEREST

REFERENCES

https://doi.org/10.17993/3cemp.2023.120151.207-224

209

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

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

1. INTRODUCTION

Mobile payment refers to a network payment method in which users use mobile

smart devices to complete transaction payments. According to different supporting

technologies, mobile payment is roughly divided into two modes: remote payment and

near-field payment. Common payment methods include online banking payment and

QR code payment [1-2]. At present, the two major characteristics of mobile payment

are convenience and security, among which the core factor affecting the development

of mobile payment business is security [3-5]. Mobile intelligent terminals are usually

bound by software and hardware disadvantages such as computing power and

storage capacity, which lead to great restrictions on the data processing capacity of

the terminal equipment [6]. Therefore, most mobile application software programs

have to implement the corresponding functions of the client by invoking remote cloud

services [7]. As far as mobile banking is concerned, the mobile banking APP of the

mobile client must complete the application functions such as transfer and wealth

management by calling the cloud service of mobile banking remotely [8]. Usually, in

the process of invoking cloud services from a mobile application, the remote

application server will first verify the identity of the end user and even the legitimacy of

the device to ensure the security of the transaction process.

In some developed countries abroad, such as Japan, South Korea, Europe and the

United States, the development of NFC payment is very rapid [9]. The mobile

payment service provided by NTTDoCoM, the largest mobile communication operator

in Japan, uses the Felica technology developed by Scmy, which uses mobile phones

with Felica chips. Its virtual wallet has contactless payment functions such as ticket

purchase and bus ride [10]. At the beginning of the promotion of mobile wallets,

NTTDoCoM installed NFC card readers for merchants for free, and made profits in the

form of monthly rent [11]. In the same year, NTTDoCoM acquired one-sixth of the

shares of Sumitomo's credit card business, so that the virtual wallet can be bound

one-to-one with bank cards. Bank card payment [12]. In 2006, the company extended

the NFC mobile payment service to the consumer credit field and launched the DCMX

mobile credit card. It can be said that mobile phones and wallets have been equated

in Japan, and shops of all sizes support NFC mobile payment [13]. In Europe, Nokia

and Philips are the main leaders of NFC mobile payment. The unified currency model

in the euro area has removed many obstacles to mobile payment. Some operators

have independent bank-authorized payment authority, which is very beneficial for

banking services, making Europe's near-field communication technology research and

development ahead of the world [14]. The first NFC experiment in Europe was

launched in March 2005 at the Frankfurt Metro, where passengers can use a Nokia

3220 mobile phone equipped with an NFC module to purchase tickets at subway

stations. The world's first multi-application NFC experiment began in October 2005.

Smartphones embedded with Philips' NFC chips were distributed to hundreds of

French citizens of Ona participating in the experiment. During the experiment, they

can be used in specific shopping malls, hotels and cinemas. After swiping the phone

to pay, the experiment went on for six months. In Munich, Germany, local residents

can use mobile phones equipped with NFC chips to swipe their cards to travel or enter

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

210

tourist attractions[15]. After Google released HCE technology in 2013, foreign markets

responded quickly: Spanish bank Bankmeter was the first to announce HCE support

earlier in 2014: TH coffee shop offered HCE-based NFC payments, and the coffee

shop invested People praised that the American coffee chain SimplyTapp also

received a large amount of investment to study the development of HCE and set out

to standardize HCE [16]. In the field of card organization, Royal Bank of Canada has

put the development of HCE on the agenda, Zaputal One of North America and

Barclays of the United Kingdom have all adjusted and applied HCE technology. Apple

launched the Apple Pay service in IOS8, which supports the NFC+AppleID transaction

method, which greatly promoted the development of NFC payment [17-18].

This paper proposes a hybrid model based on Bilateral AM (BAM) and CNN-GRU

(Gated Recurrent Unit) to make multivariate predictions on the blockchain, an

important factor influencing mobile payment, and selects the same as the blockchain.

Other relevant feature performance parameters are used as input [19-20] and the

blockchain is used as output. First, a feature AM is introduced on the input side to

quantify the relationship between performance parameters and blockchain output;

then, through the powerful feature extraction capability of one-dimensional (One

Dimensional, 1D-CNN), the local information between the input information is mined.

Finally, a time-series feature AM is introduced on the output side to strengthen the

expression of important information at historical moments for the prediction output

[21]. The purpose is to establish a prediction model that depends on each

performance parameter under the time node, explore the connection between mobile

payment and blockchain, and accurately predict the impact of blockchain on mobile

payment.

2. PRINCIPLES OF DEEP LEARNING MODELS

2.1. 1D-CNN

The convolutional neural network performs high-dimensional feature mapping on

the original data through local connection and weight sharing, and mines the feature

information of the original data. 1D-CNN is mainly used to process time series, and its

internal structure is shown in Figure 1. For processing time series, the convolution

layer extracts the translation features of the data in the direction, and extracts the

effective feature vectors on the time series. From an accurate point of view, the

analysis is to perform cyclic product and summation on the data. The specific

expression is as follows:

where y, w, v are sequences, µ is the number of convolutions, and M is the length

of v.

(1)

( ) ( ) ( ) ( ) ( )

y w v w v M

µ µ µ µ τ τ

= = −

∑

https://doi.org/10.17993/3cemp.2023.120151.207-224

211

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

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

Fig.1 1D-CNN

2.2. ATTENTION MECHANISM (AM)

Attention mechanism (AM)is a reserve apportionment model that pretends the

attention of the human brain. The human brain can pay attention to some important

information of interest and ignore irrelevant information at a certain time node. The

attention distribution of different information can be more important. The attention

model assigns greater weight to important information through this probability

distribution mode to different information, thereby improving the model's extraction of

important information and improving the prediction accuracy of the model. In time

series prediction, the AM can not only act on the input side to reflect the degree of

correlation between each feature parameter and the predicted output, but also on the

output side to weight the information at the historical moment to highlight the

information related to the current prediction. Important time point information [22].

2.3. GRU

GRU network is an improved mode of LSTM network [23]. By optimizing the gate

structure inside the LSTM network, the input gate and the forget gate are combined

into an update gate, and the state of the neuron is mixed with the state of the hidden

layer. The update gate inputs the combined matrix of the input vector and Xt

and the

state memory variable ht - 1

of the previous moment into the update gate after the

nonlinear transformation of the activation function, and determines the degree to

which the information of the previous moment is retained to the current state [24]. The

reset gate combines the previous state information with the current state information

in the manner of 1 - zt times ht - 1 and zt times h

t

as the output of the current state

information. The GRU structure network is exposed in Fig. 2, and expression is

revealed in Equation (2) [25].

Input layer Convolution

Layer Pool layer Full

Connecrion Output layer

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

212

In Figure 2 and equation (2): Xt，ht - 1，rt，zt，h

t，ht are the input information, the

state information of the previous moment, the update gate, the reset gate, the input

vector and the previous hidden layer state information, respectively. Summary, the

output of the current hidden layer state; Wr，Wz，Wh

 are the trainable weight matrices

of each gate state; σ is represented as the Sigmoid function.

MOBILE PAYMENT AND BLOCKCHAIN PREDICTION

MODEL BASED ON BAM AND CNN-GRU HYBRID

MODEL

3.1. MOBILE PAYMENT AND BLOCKCHAIN PREDICTION

MODEL

Mobile payment is not only carefully associated the operation of various

components in the system, but also related to the blockchain environment, such as

the sharing economy, etc., but these are external factors and will indirectly affect the

operation status of each component in the system , and then make the mobile

payment security change. Therefore, according to the internal operation rules of

mobile payment, we select sharing economy N1, Internet of Things N2

, cloud

computing Wf, artificial intelligence po and digital economy To

, five performance

parameters that are highly correlated with mobile payment as input parameters, and

establish the following parameters at each time point. The relevant performance

parameters are dependent on the predictive model.

Note that the time series set of mobile payment is

, and the time series of the five blockchain-

related features of the input terminals N1, N2, Wf, po and To are X = [ x 1，x 2，...，

xT ] = [ x(1)，x(2 )，...，x(5) ]′, the specific expansion can be represented by equation

(3). Where represents a set of five related

feature parameter variables measured at time t,

is represented as the measured

value sequence of the k-th relevant feature parameter at T historical moments.

(2)

[ ]

( )

[ ]

( )

[ ]

()

( )

tanh ,

t t t t

t z t t

t t t t

r W h X

z W h X

h W r h X

h z h z h

−

⎧=⋅

⎪=⋅

⎪

⎨=⋅ ×

⎪

⎪=− × +×

⎩



Y=[y1，y2，...，yT]∈RT

xt=[xt(1)，xt(2)，...，xt(5)]

x(k)=[x1(k)，x2(k)，...，xT(k)]

https://doi.org/10.17993/3cemp.2023.120151.207-224

213

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

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

where k (1≤k≤5) of x( k ) represents the number of features. (t 1≤t≤

T) of xt

represents the value of the corresponding feature at time t.

According to the five characteristic performance parameter variables related to the

blockchain as the input, and the blockchain at the corresponding time as the output of

the model, the multi-variable prediction of mobile payment in the future time and time

is carried out. Let the mapping function of the entire model be Fθ

, then the predicted

value represents for:

In order to obtain the relationship between each feature performance parameter

and the blockchain at the current moment and the dependency in the time series

information, a bilateral attention and CNN-GRU hybrid model combining the feature

AM and the time series AM are adopted. Multivariate forecasting methods. A feature

AM is presented on the input side to calculate the degree of correlation between the

exhaust gas temperature value to be measured and other related performance

parameters, so that the features with strong correlation are assigned greater weights,

while the weak or irrelevant features are weakened. information. CNN mines the high-

dimensional features of the input information through operations such as convolution

pooling, and effectively reduces the error caused by manual feature extraction. The

time series AM is introduced at the output, independently select the information of the

historical key moment with high correlation with the current moment, and solve the

problem of GRU network for long-term sequence.

3.2. FEATURE AM

In order to obtain the degree of correlation between the five feature parameters and

the blockchain to be tested, a feature AM is introduced on the input side, and the

multi-layer perceptron calculation method is used to quantify the attention weights of

various features. The model is shown in Figure 2 [26]. 

(3)

( ) ( ) ( )

1 2 5

1 1 1

1 2 5

T 5

2 2 2

1 2 5

T T T

x x x

⎡ ⎤

⎢ ⎥

=∈

⎢ ⎥

⎣ ⎦







X R

(4)

( )

1 1 2

, , ,

T T

y F x x x

θ+

= …



https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

214

Fig. 2 Characteristic attention model

The five feature parameters at time t are combined with the hidden layer state ht - 1

at the preceding time as the input of the feature AM. The weight of each feature

parameter at this time is calculated by formula (5), and the Softmax function is used.

Normalize e( k ) t according to formula (6), namely:

where Ve, We and Ue

are the weight matrices of the feature AM, and be is the

corresponding bias term.

The feature parameter weight λ

( k ) t assigned according to the feature AM is

multiplied by the corresponding feature input x( k ) t to obtain the associated features

x with different feature contribution rates, so as to perform strong and weak correlation

for each feature Different expressions can be specifically expressed as:

Finally, the input information x



is iterated by formula (8) to ensure that the hidden

layer state ht at each time t contains the associated feature x:

where fGRU1 represents the network unit of the input side GRU.

3.3. CNN LAYER

The introduction of the 1D-CNN network is to extract the feature of the relationship

information processed by the feature AM, map the relationship information to the high-

dimensional feature space, mine the deep-level feature information, and extract the

GRU

Feature weight

(1)

(2)

(5)

i(5)

(2)

i(1)

(2)

(5)

t-1

Characteristic attention

mechanism x

i(1)

i(2)

i(5)

(1)

(2)

i(5)

Sofrmax

(5)

(6)

( ) ( )

( )

e e 1 e e

relu

k k

t t

e x

−

= + +V W h U b

( )

exp

=∑

(7)

( ) ( ) ( ) ( ) ( ) ( )

( )

1 1 2 2 5 5

, , ,

t t t t t t t

x x xλ λ λ= …



(8)

( )

GRU1 1

t t t

−

=

h h x

https://doi.org/10.17993/3cemp.2023.120151.207-224

215

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

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

key node information in the feature variables [27]. The convolutional layer extracts

features, the pooling layer filters the information, and the dropout layer discards some

neurons to prevent the network from over-fitting due to over-reliance on some local

features [28]. In this paper, at the connection between 1D-CNN and GRU, the

maximum pooling layer and dropout layer are used to replace the fully connected

layer. This operation not only reduces the data dimension input to the GRU network,

reduces the network training time, but also preserves the time series information of

the input features to the greatest extent, ensuring the prediction accuracy of the

model. The output feature vector Rφ of the 1D-CNN network can be expressed as:

where H is the set of hidden layers of the input side GRU, the outputs C and P are

the outputs of the convolutional layer and the pooling layer, respectively; W and b1 are

the weights and bias terms of the convolutional layer; b2

is the bias of the pooling

layer. set item; the output of the CNN layer is:

3.4. TEMPORAL AM

Since the predicted value of the blockchain is greatly affected by the historical

state, and the hidden layer state information at different times has different effects on

the output of the current network, in addition, the network output is more inaccurate

due to the increase in the length of the time series. In order to enable the predicted

value to process historical state information autonomously and to strengthen the

expression of important historical moment information with high output relevance at

the current moment, this paper introduces a time-series AM for the output side of the

GRU network. The specific structure is shown in Figure 4 [29].

(9)

(10)

( )

relu=⊗+C H W b

( ) 2

maxpolling

= +P C b

(11)

1 2 t ,

, , , , ,r r r r

ϕ ϕ

⎡ ⎤

= … …

⎣ ⎦

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

216

Fig.4 Temporal attention model

The time series information of the hidden layer state including the relationship

information processed by the CNN is the vector Rφ

, which is used as the input of the

GRU network, and its output is expressed as C, and the output at time t is expressed

as:

where fGRU2 represents the network unit of the output side GRU.

The input of the time-series AM is the output vector C processed by the GRU

network on the output side. According to the AM, the historical state information at

each time point is weighted and expressed, and the Softmax function is continued to

normalize the weight β

tT, and record each moment in history. The correlation degree

of the hidden layer state information to the output at the current moment is αφt [30].

where Vc and Wc

are the corresponding weight matrices of time-series attention,

and bc is the bias.

t indicates that the correlation degree of each hidden layer state information in the

history to the prediction output at the current time is quantified, and all α φ

t and the

corresponding hidden layer state information are weighted and summed, and the

output of the time series AM at time t is lt express.

GRU GRU GRU GRU

Tempporal weight

Softmax

c1c2ci-1

α(1)

βi(2) βi(r-1) βi

(r)

α(2) α(r-1) α(r)

...

Output layer

βi(1)

Temporal attention

mcchanism

...

(12)

( )

GRU2 1

t t t

c f c r

−

(13)

(14)

( )

exp

=∑

( )

c c c

relu

t t

c bβ= +V W

https://doi.org/10.17993/3cemp.2023.120151.207-224

217

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

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

Finally, the output information is dimensionally transformed through the fully

connected layer network, and the final blockchain prediction value yT + 1 is obtained:

where Wy and by

are the weights and biases for dimensional changes in the

network.

3.5. HYBRID MODEL BASED ON BAM AND CNN-GRU

The structure of the entire prediction model is shown in Figure 5. After the input

information is processed by the feature AM, relevant information with different weights

is obtained, and then it enters the GRU network for learning. The output of the

network is used as the input of 1D-CNN. In the processing of the pooling layer and

dropout layer, the information enters the GRU network on the output side, and the

hidden layer state is used as the input of the timing AM on the output side. . In the

training process of this model, the Adam (Adaptive Moment Estimation) optimization

algorithm is selected to update and learn various parameters, and the loss function of

the model adopts the mse function, as shown in formula (17) [31].

In the formula, n is the number of samples; yi is the actual value of the blockchain,

and yi is the blockchain value predicted by the model.

Fig.5 Model structure based on BAM and CNN-GRU

3.6. ERROR ANALYSIS

In this paper, Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and

Mean Absolute Percentage Error (MAPE) are used as indicators to evaluate the

prediction accuracy of each model. The expressions are as follows shown [32]:

(15)

t t t

l cα

=∑

(16)

( )

1 1 2

, , , tanh

T T y t y

y F x x x W l b

θ+= … = +



(17)

( )2

i i

Loss y y

=−

∑

Ounput layer

Temporal attcntion necttananism

IDCNN

Characterastac attention mcchartism

Dcopout

...

GRU GRU GRU

...

i-1

GRU GRU GRU

...

Comolution

layer

Dcopout

Maxpcoling

layer

Input layer

t-1

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

218

4. MODEL PREDICTIONS

For the sample data, the mixed model of BAM and CNN-GRU, BAM-GRU model,

CNN-GRU model, GRU model, LSTM model, BP model and SVM model proposed in

this paper are used to predict the output. The training and test sets are split in a ratio

of 4:1. The GRU network and LSTM network structure in the above models all use the

same hyperparameters (the number of hidden layer neurons is 128, the sliding

window is 6, the number of iterations is 400, and the batch_size is 256); the AM uses

relu as the internal structure[33-34]. Activation function, and use Softmax function to

normalize it; the convolution layer of CNN network is set to 64, the convolution kernel

is 1, the maximum pooling layer is 4, and the dropout is 0.3; BP network adopts the

number of hidden layer neurons is a network structure of 8; the SVM network adopts

radial basis kernel function.

According to the three evaluation indicators selected in this paper, the prediction

performance and accuracy of different models 566 are evaluated. The experimental

comparison results are shown in Table 1.

Table 1 Comparison of prediction accuracy of different models

From the information in Table 1, it can be concluded that the prediction accuracy of

the algorithm in this paper is better than that of the other 6 algorithms. Compared with

the other 6 methods, MAE is reduced by 75.45%, 64.74%, 62.84%, 59.41%, 45.54%,

44.16%; RMSE Compared with the other 6 methods, the reductions were 60.39%,

51.86%, 47.84%, 42.90%, 39.26%, 24.27%, respectively; compared with the other 6

methods, MAPE decreased by 0.50%, 0.31%, 0.29%, 0.25%, 0.15%, 0.13%.

(18)

(19)

(20)

1100%

ni i

y y

MAPE n y

−

=×

∑



i i

MAE y y

=−

∑

( )2

i i

RMSE y y

=−

∑



Model AE RMSE MAPE

BP 4.48 4.57 0.67

SVM 3.12 3.76 0.48

LSTM 2.96 3.47 0.46

GRU 2.71 3.17 0.42

CNN-GRU 2.02 2.98 0.32

BAM-GRU 1.97 2.39 0.30

Proposed 1.10 1.81 0.17

https://doi.org/10.17993/3cemp.2023.120151.207-224

219

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

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

Comprehensive analysis, the algorithm in this paper has obvious reduction in the

three error evaluation indicators, indicating that the algorithm in this paper has a

relatively good prediction performance. According to the analysis of the error results,

the machine learning methods (BP, SVM) are not as outstanding as the deep learning

methods (LSTM, GRU, CNN-GRU, BAM-GRU, Proposed) in terms of prediction

effect. From Table 1, it can be seen that the CNN-GRU model and the BAM-GRU

model have a significant decrease in the three error evaluation indicators compared

with the single deep learning model, indicating that the CNN network can be

performed by the high-dimensional local dependence on the multivariate input

performance parameters. Mining to improve the prediction performance of the model.

By introducing BAM, the degree of correlation between input features and output is

quantified on the input side, and the contribution rate of each performance feature is

adaptively extracted, which effectively avoids the output expression of non-critical

information and secondary feature information, and strengthens important information

for prediction. The output side strengthens the correlation expression of historical

important information for the current prediction output, reduces information omission

and memory decay, and solves the prediction lag problem of LSTM and GRU single

network models.

The comparison of the prediction output curves of each model on the test set is

shown in Figure 6 and Figure 7. It can be seen from Figure 6 that the traditional

machine learning method has a poor prediction effect, and the predicted output value

has a low degree of fitting with the actual value. The error is large.

Fig. 6 Comparison of predicted values between the proposed model and the

machine learning model

Figure 7 shows that the prediction output of the deep learning model has a high

degree of fitting with the actual value, which proves the advantages of deep learning

in time series prediction. The hybrid model of BAM and CNN-GRU proposed in this

paper can not only accurately predict in a slightly smooth interval, but also accurately

capture the changing law of mobile payment in high and low peak time periods. The

other four learning methods can also accurately predict the blockchain in some

intervals, but there is still a certain gap between their prediction performance and the

method proposed in this paper when the blockchain peaks and fluctuates violently. It

shows that the model proposed in this paper has good performance in establishing

long-term dependencies of time series and effectively capturing the dynamic changes

0 100 200 300 400 500

610

620

630

640

650

660

670

Numerical value

Sample

Proposed

Ture

SVM

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

220

of blockchain. By accurately predicting the blockchain, the change rule of the

blockchain can be known in advance, and compared with the corresponding baseline

value to check whether the difference is within the maximum range allowed for

operation, and then take the corresponding maintenance strategy. In addition, when it

is found that there is a sudden and substantial rise and fall of the blockchain within a

certain predicted time period, it is necessary to consider whether the mobile payment

operation is normal, and investigate the reasons in time to avoid security accidents.

Fig. 7 Comparison between the predicted value of the proposed model and the

deep learning model

5. CONCLUSION

Aiming at the relationship between mobile payment and blockchain economy, this

paper proposes a hybrid model based on BAM and CNN-GRU to improve the

prediction accuracy and stability of the long mobile payment model. The following

conclusions are drawn: (1) The prediction accuracy of the hybrid model of BAM and

CNN-GRU proposed in this paper is better than that of the other six algorithms.

Compared with the other six methods, the MAE is reduced by 75.45%, 64.74%,

62.84%, respectively. 59.41%, 45.54%, 44.16%; (2) The hybrid model of BAM and

CNN-GRU proposed in this paper can not only accurately predict in a slightly smooth

interval, but also accurately capture the change law of mobile payment in high and low

peak periods; ( 3) The CNN network can improve the prediction performance of the

model by mining the high-dimensional local dependencies of multivariate input

performance parameters.

6. CONFLICT OF INTEREST

The authors declared that there is no conflict of interest.

0 100 200 300 400 500

610

620

630

640

650

660

670

Numerical value

Sample

Proposed

Ture

LSTM

GRU

GRCNN-GRU

BAM-GRU

https://doi.org/10.17993/3cemp.2023.120151.207-224

221

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

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

REFERENCES

(1) Ayu, W., Siti, S., & Muhardono, A. (2021).

Consumer Behavior Toward

Adoption of Mobile Payment: A Case Study in Indonesia During the

COVID-19 Pandemic. Journal of Asian Finance Economics and Business(4).

https://doi.org/10.13106/JAFEB.2021.VOL8.NO4.0581

(2) Yang, S., Lu, Y., Gupta, S., Cao, Y., & Zhang, R. (2012).

Mobile payment

services adoption across time: An empirical study of the effects of

behavioral beliefs, social influences, and personal traits.

Computers in

Human Behavior, 28(1), 129-142. https://doi.org/10.1016/j.chb.2011.08.019

(3) Fang, C. C., Liou, J., Huang, S. W., Wang, Y. C., & Tzeng, G. H. (2021). A

Hybrid, Data-Driven Causality Exploration Method for Exploring the Key

Factors Affecting Mobile Payment Usage Intention. https://doi.org/10.3390/

math9111185

(4) Tao, & Zhou. (2013). An empirical examination of continuance intention of

mobile payment service. Decision Support Systems. https://doi.org/10.1016/

j.dss.2012.10.034

(5) Zhou, T. (2011). An Empirical Examination of Initial Trust in Mobile Payment.

Internet Research Electronic Networking Applications & Policy

, 77(2),

1519-1531. https://doi.org/10.1108/10662241111176353

(6)

Chakraborty, D., Siddiqui, A., Siddiqui, M., Rana, N. P., Dash, G., &

Timmermans, H. (2022). Mobile payment apps filling value gaps: Integrating

consumption values with initial trust and customer involvement. Journal of

Retailing and Consumer Services, 6 6. https://doi.org/10.1016/

j.jretconser.2022.102946

(7)

Handarkha, Y. D., Harjoseputro, Y., Samodra, J. E., & Irianto, A. (2021).

Understanding proximity mobile payment continuance usage in Indonesia

from a habit perspective. Journal of Asia Business Studies. https://doi.org/

10.1108/JABS-02-2020-0046

(8) Zhao, Y., & Bacao, F. (2021).

How Does the Pandemic Facilitate Mobile

Payment? An Investigation on Users' Perspective under the COVID-19

Pandemic. International Journal of Environmental Research and Public Health,

18(3), 1016.https://doi.org/10.3390/ijerph18031016

(9) Fan, L., Zhang, X., Rai, L., & Du, Y. (2021).

Mobile Payment: The Next

Frontier of Payment Systems? - an Empirical Study Based on Push-Pull-

Mooring Framework. Journal of Theoretical and Applied Electronic Commerce

Research, 16(2), 179-193. https://doi.org/10.4067/S0718-18762021000200112

(10) Becker, Michel, Stolper, Oscar and Walter, Andreas.

What Drives Mobile

Banking Adoption? – An Empirical Investigation Using Transaction Data

Zeitschrift für Bankrecht und Bankwirtschaft

, vol. 34, no. 1, 2022, pp. 1-11.

https://doi.org/10.15375/zbb-2022-0103

(11) He, X., Gupta, S., Hu, Y., & Zhao, L. (2021). Understanding the dual role of

habit in cross-channel context: an empirical analysis of mobile payment.

International Journal of Mobile Communications, 19(1), 1. https://doi.org/

10.1504/IJMC.2021.10034125

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

222

(12) Kang, D. (2021).

Research on Implicit Attitude and Behavioral Tendency of

College Students to Mobile Payment. Psychology of China

, 3(2), 103-110.

https://doi.org/10.35534/pc.0302014

(13)

Geomagnetic secular variation anomalies in relation to the recent crustal

movement in the southwestern region of Japan: Sumitomo, Norihiko, 1981

Bull. Disaster Prevent. Res. Inst. Kyoto Univ., 30(4)(274):97–130,

Deep Sea

Research Part B. Oceanographic Literature Review

, Volume 28, Issue 12, 1981,

Page 874, ISSN 0198-0254, https://doi.org/10.1016/0198-0254(81)91240-1

(14) Kim, H. R., Zhou, W., & Yun, S. (2021).

Mobile Payment Service Continuance

Usage Intention and Network Externality: Focused on Self and Functional

Congruity. THE JOURNAL OF SOCIAL SCIENCE, 28(1), 176-197. https://

doi.org/10.46415/jss.2021.03.28.1.176

(15) Ramli, F., & Hamzah, M. I. (2021).

Mobile payment and e-wallet adoption in

emerging economies: A systematic literature review.

Journal of Emerging

Economies and Islamic Research, 9(2), 1. https://doi.org/10.24191/

jeeir.v9i2.13617

(16) Wong, D., Liu, H., Yue, M. L., Sun, Y., & Zhang, Y. (2021).

Gamified Money:

Exploring the Effectiveness of Gamification in Mobile Payment Adoption

among the Silver Generation in China.

Information Technology & People,

ahead-of-print(ahead-of-print). https://doi.org/10.1108/ITP-09-2019-0456

(17) Cacioppo, J. T., Petty, R. E., & Kao, C. F. (1984).

The efficient assessment of

NFC. Journal of Personality Assessment, 48(3), 306-307. https://doi.org/

10.1207/s15327752jpa4803_13

(18) Sehaqui, H., Qi, Z., & Berglund, L. A. (2011).

High-porosity aerogels of high

specific surface area prepared from nanofibrillated cellulose (NFC)

Composites Science & Technology, 71(13), 1593-1599. https://doi.org/10.1016/

j.compscitech.2011.07.003

(19) Christidis, K., & Devetsikiotis, M. (2016).

Blockchains and Smart Contracts for

the Internet of Things. IEEE Access, 4, 2292-2303. https://doi.org/10.1109/

ACCESS.2016.2566339

(20) Yli-Huumo, J., Ko, D., Choi, S., Park, S., & Smolander, K. (2016).

Where Is

Current Research on Blockchain Technology?—A Systematic Review

PLoS ONE, 11(10), e0163477. https://doi.org/10.1371/journal.pone.0163477

(21) Wang, Y., Yan, J., Yang, Z., Wang, J., & Geng, Y. (2021).

A novel 1DCNN and

domain adversarial transfer strategy for small sample GIS partial discharge

pattern recognition. Measurement Science and Technology

, 32(12), 125118

(125110pp). https://doi.org/10.1088/1361-6501/ac27e8

(22) Palomino, A. J., Marfil, R., Bandera, J. P., & Bandera, A. (2011).

A Novel

Biologically Inspired AM for a Social Robot.

Eurasip Journal on Advances in

Signal Processing, 2011(1), 1-10. https://doi.org/10.1155/2011/841078

(23)

Jiang, Y., Zhao, M., Zhao, W., Qin, H., Qi, H., Wang, K., & Wang, C. (2021).

Prediction of sea temperature using temporal convolutional network and

LSTM-GRU network. Complex Engineering Systems, 1(2), -. https://doi.org/

10.20517/ces.2021.03

(24) Huang, G., Li, X., Zhang, B., & Ren, J. (2021

). PM2.5 Concentration

Forecasting at Surface Monitoring Sites Using GRU Neural Network Based

https://doi.org/10.17993/3cemp.2023.120151.207-224

223

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

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

on Empirical Mode Decomposition. Science of The Total Environment

, 768(3),

144516. https://doi.org/10.1016/j.scitotenv.2020.144516

(25) Xu, S., Li, J., Liu, K., & Wu, L. (2019).

A Parallel GRU Recurrent Network

Model and its Application to Multi-Channel Time-Varying Signal

Classification. IEEE Access, PP(99), 1-1. https://doi.org/10.1109/

ACCESS.2019.2936516

(26) Shen, X., Shao, Z., & Tian, Y. (2014).

Built-Up Areas Extraction by Textural

Feature and Visual AM. Acta Geodaetica et Cartographica Sinica. https://

doi.org/10.13485/j.cnki.11-2089.2014.0131

(27) Razavian, A. S., Azizpour, H., Sullivan, J., & Carlsson, S. (2014). C

NN Features

off-the-shelf: an Astounding Baseline for Recognition. IEEE. https://doi.org/

10.1109/CVPRW.2014.131

(28) Yan, C., Pang, G., Bai, X., Liu, C., Xin, N., Gu, L., & Zhou, J. (2021).

Beyond

triplet loss: person re-identification with fine-grained difference-aware

pairwise loss. IEEE Transactions on Multimedia. https://doi.org/10.1109/

TMM.2021.3069562

(29) Ning, X., Gong, K., Li, W., & Zhang, L. (2021).

JWSAA: joint weak saliency

and attention aware for person re-identification. Neurocomputing

, 453,

801-811. https://doi.org/10.1016/j.neucom.2020.05.106

(30)

Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun

Zhou, Edwin Hancock,

Uncertainty estimation for stereo matching based on

evidential deep learning, Pattern Recognition

, Volume 124, 2022, 108498,

ISSN 0031-3203, https://doi.org/10.1016/j.patcog.2021.108498.

(31)

Liu Ying, Zhang Qian Nan, Wang Fu Ping, Chiew Tuan Kiang, Lim Keng Pang,

Zhang Heng Chang, Chao Lu, Lu Guo Jun & Ling Nam.

Adaptive weights

learning in CNN feature fusion for crime scene investigation image

classification. Connection Science.

https://doi.org/

10.1080/09540091.2021.1875987

(32) Yu, Z., Li, S., Sun, L. N. U., Liu, L., & Haining, W.

Multi-distribution noise

quantisation: an extreme compression scheme for transformer according

to parameter distribution. https://doi.org/10.1080/09540091.2021.2024510

(33) Dewani, A., Memon, M. A., y Bhatti, S. (2021).

Development of computational

linguistic resources for automated detection of textual cyberbullying

threats in Roman Urdu language.

3C TIC. Cuadernos de desarrollo aplicados

a las TIC, 10(2), 101-121. https://doi.org/10.17993/3ctic.2021.102.101-121

(34) Guan Fuyu,Cao Jie,Ren Jie & Song Wenli.(2021).

The teaching of sports

science of track and field-based on nonlinear mathematical equations

Applied Mathematics and Nonlinear Sciences(1). https://doi.org/10.2478/

AMNS.2021.2.00155.

https://doi.org/10.17993/3cemp.2023.120151.207-224

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

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

224