Research on classification of e-commerce customers based on BP neural network

The customers’ demands are infinite, and the limited resources owned by enterprises cannot meet the needs of all customers. In order to maximise social benefits as much as possible, enterprises must make rational use of limited resources and strive for high-value customers, which requires enterprises to have the ability to judge the value of customers. By implementing market segmentation by means of technology, enterprises can identify customers who can promote the development and then focus the limited resources on these potential customers [6]. According to the existing research, it is found that the measurement indexes of customer classification are diverse and the process is complicated, which mainly includes the following points: Firstly, enterprises should build an index system of customer classification. Different products and services of enterprises determine the difference in selected indicators. Therefore, enterprises should choose appropriate indicators according to the characteristics of their target customers [7]. At present, customer consumption involves a lot of information, including not only the basic attribute information of customers, such as customer gender, age and education level, but also other information of customer consumption, such as consumption frequency, total consumption amount and consumption satisfaction [8]. Secondly, enterprises should use big data to sort out the collected data, or other information technology methods to process the data; understand the customer consumption characteristics in a certain way; and calculate the customer value quantitatively [9]. Thirdly, according to the calculated customer value, enterprises should classify and manage customers and put forward appropriate suggestions for each type of customers to obtain maximum benefits [10]. The process of customer classification is shown in Figure 1.

Fig. 1

Compared with the characteristics of customer classification in the traditional business environment, customer classification in the e-commerce environment has changed a lot: (1)

Classification indicators: Classification indicators are diversified. In the environment of e-commerce, enterprises classify customers not only by relying on customer value or some basic characteristics of customers but also by adding some log files left by customers on the Internet, from which personalised factors such as customers’ use feelings, psychological changes, preferences and habits can be analysed. These indicators can more realistically reflect the real needs of customers [11].

Choosing appropriate indicators is the key to the quality of customer classification results. A good classification indicator system should follow the following principles.

(1)

The scientific principle: The selection of indicators must have a theoretical basis, and indicators without practical significance cannot be selected. Generally speaking, the selection of indicators should follow the combination of science and practice, dynamic and static, and qualitative and quantitative [16].

Generally, selecting the customer classification index in the e-commerce environment is an extremely critical and complicated process. In order to ensure the effective development of classification, it is necessary to combine the particularity and interactivity of enterprises to construct a classification indicator suitable for the actual situation of enterprises [18–22].

This work will build a customer classification index system from the following three dimensions: characteristics of e-commerce customers, current values and potential values of the customers. The index system established from the aforementioned dimensions can also calculate the value of e-commerce customers more comprehensively and scientifically, as shown in Figure 2.

Fig. 2

(1)

Characteristics of e-commerce customer

According to the existing research and practical basis, it can be found that the characteristics of e-commerce customers mainly include six secondary indicators: age, income, education level, geographical location, trust and customer interaction value. At the age of 26–40 years, the scale of monthly active users of Internet is the largest, and according to statistical data, it is also found that the middle-aged group is the main service target of e-commerce [23]. From the perspective of education level, the proportion of Internet users with junior high school education is the largest, followed by senior high school/technical secondary school education. As e-commerce is a new thing, the higher the education level, the greater the acceptance, so it can be considered that the customer value of those with a high education level is greater. From the perspective of income, Internet users with a monthly personal income of 3,001–5,000 yuan account for the largest proportion in China. Secondly, for the network users with 5,001–8,000 yuan and above 8,000 yuan, the consumption of e-commerce products especially needs the support of income, so it can be considered that the customer value above 8,000 yuan is relatively the largest [24]. Enhancing customers’ trust in the e-commerce platform and regional advantages and bringing good shopping experience to customers are conducive to enhance customers’ aspiration to consume. Therefore, it can be considered that customers with better geographical location have greater value.

The e-commerce customer classification index system is constructed from the customer characteristics, current value and potential value, which comprehensively considers all factors that may affect the customer value, and it is scientific to calculate the customer value with this index system. According to the existing literature research, realistic statistical data and cases, an e-commerce customer classification index system is established, which is shown in Table 1.

(1)

Weight setting

The three indicators in the current customer value reflect the different importance of customers to the enterprise. Accurate determination of the weight of each specific indicator plays an important role in customer pre-classification in the current value. At present, the methods to determine the index weight include expert grading method and analytic hierarchy process. In this study, the expert grading method is used to establish the importance of the index, and the corresponding weight set is a={a1,a2,⋯,an} , and Σai=1 . After the index weight is determined, the customer’s current value can be calculated according to the corresponding index data of each sample, as shown in Figure 3.

Fig. 3

The BP neural network is a multi-layer feedforward neural network, which consists of input layer, hidden layer and output layer [28]. The learning process of the network consists of forward propagation and backward propagation. In the forward propagation, the data pass through the input layer, the hidden layer and the output layer in turn. If the result fails to meet the expected requirements, it will enter the backward propagation, where it distributes the error to each small unit in the network structure, and through constant adjustment, until the output result reaches the expectation [29].

When calculating each node, it needs to use the S-type function for basic calculation, so it requires a hidden layer to solve the problem of decision classification. When there are two hidden layers, the output function can be used on the input image at will. However, for a small network, one hidden layer is enough to complete the work [30].

The actual output is calculated in the direction from input to output, while the weights and thresholds are corrected in the direction from output to input [31].

x_j represents the input of the jth node of the input layer, j=1,.,m ;

w_ij represents the weight between the ith node of the hidden layer and the jth node of the input layer;

θ_i represents the threshold of the ith node of the hidden layer;

ϕ(x) represents the excitation function of the hidden layer;

w_ki represents the weight between the kth node of the output layer and the ith node of the hidden layer, with i=1,.,q ;

a_k represents the threshold of the kth node of the output layer, k=1,.,l ;

Ψ(x) represents the excitation function of the output layer; and

O_k represents the output of the kth node of the output layer.

(1) Forward propagation of the signal

Input net_i of the ith node of the hidden layer: 2 neti=∑j=1Mwijxj+θi

Output y of the ith node of the hidden layer_i: 3 yi=ϕ(neti)=ϕ(∑j=1Mwijxj+θi) $${y_i} = \phi (ne{t_i}) = \phi \left( {\mathop \sum \limits_{j = 1}^M {w_{ij}}{x_j} + {\theta _i}} \right)$$

Output the input net of the kth node of the layer_k: 4 netk=∑i=1qwkiyi+ak=∑i=1qwkiϕ(∑j=1Mwijxj+θi)+ak $$ne{t_k} = \mathop \sum \limits_{i = 1}^q {w_{ki}}{y_i} + {a_k} = \mathop \sum \limits_{i = 1}^q {w_{ki}}\phi \left( {\mathop \sum \limits_{j = 1}^M {w_{ij}}{x_j} + {\theta _i}} \right) + {a_k}$$

Output o of the kth node of the output layer: 5 ok=ψ(netk)=ψ(∑i=1qwkiyi+ak)=ψ(∑i=1qwkiϕ(∑j=1Mwijxj+θi)+ak) $${o_k} = \psi (ne{t_k}) = \psi \left( {\mathop \sum \limits_{i = 1}^q {w_{ki}}{y_i} + {a_k}} \right) = \psi \left( {\mathop \sum \limits_{i = 1}^q {w_{ki}}\phi \left( {\mathop \sum \limits_{j = 1}^M {w_{ij}}{x_j} + {\theta _i}} \right) + {a_k}} \right)$$

(2) Backward propagation of error

Backward propagation of error, that is, firstly, the output error of neurons in each layer is calculated layer by layer from the output layer, and then the weights and thresholds of each layer are adjusted according to the error gradient descent method, so the final output of the modified network can approach the expected value.

The quadratic error criterion function for each sample P is E_p: 6 Ep=12∑k=1L(Tk−ok)2

The total error criterion function of the system for P training samples is 7 E=12∑p=1P∑k=1L(Tkp−okp)2

According to the error gradient descent method, the output layer weight correction Δw_ki, output layer threshold correction Δa_k, hidden layer weight correction Δw_ij and hidden layer threshold correction Δθ_i are considered as follows: 8 Δwki=−η∂E∂wki;Δak=−η∂E∂ak;Δwij=−η∂E∂wij;Δθi=−η∂E∂θi $${\rm{\Delta }}{w_{ki}} = - \eta {{\partial E} \over {\partial {w_{ki}}}};\>{\rm{\Delta }}{a_k} = - \eta {{\partial E} \over {\partial {a_k}}};\>{\rm{\Delta }}{w_{ij}} = - \eta {{\partial E} \over {\partial {w_{ij}}}};\>{\rm{\Delta }}{\theta _i} = - \eta {{\partial E} \over {\partial {\theta _i}}}$$

The weight adjustment formula of the output layer: 9 Δwki=−η∂E∂wki=−η∂E∂netk∂netk∂wki=−η∂E∂ok∂ok∂netk∂netk∂wki

The output layer threshold adjustment formula: 10 Δak=−η∂E∂ak=−η∂E∂netk∂netk∂ak=−η∂E∂ok∂ok∂netk∂netk∂ak

The weight adjustment formula of the hidden layer: 11 Δwij=−η∂E∂wij=−η∂E∂neti∂neti∂wij=−η∂E∂yi∂yi∂neti∂neti∂wij

The hidden layer threshold adjustment formula: 12 13 14 $$\matrix{ {{{\partial ne{t_k}} \over {\partial {w_{ki}}}}} & { = {y_i},\>{{\partial ne{t_k}} \over {\partial {a_k}}} = 1,\>{{\partial ne{t_i}} \over {\partial {w_{ij}}}} = {x_j},{{\partial ne{t_i}} \over {\partial {\theta _i}}} = 1} \cr } $$ 15 16 17

The following formula is finally obtained: 18 19 20 21

The BP network is a one-way multi-layer feedforward network, which includes input layer, hidden layer and output layer. It is a widely used model at present. The algorithm adopts error backward propagation learning method in the hierarchical network structure, and the learning process consists of forward propagation and error back propagation.

The algorithm of the BP neural network is a kind of learning algorithm that has the characteristics of supervision. The main idea is that for q input learning samples P1, P2,., PQ, the expected output samples corresponding to them are known as T1, T2,., TQ. The learning algorithm trains the algorithm and corrects the weights by comparing the actual outputs A1, A2, ⋯⋯ of the network with the gaps between the targets T1, T2,., TQ, which makes A and T as close as possible (Figure 4).

Fig. 4

The three-layer network structure is often used in the research of customer classification in the BP neural network algorithm, so three-layer BP neural network structure is selected to calculate customer value, and the preparatory work includes determining the number of network layers, the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes and data processing.

4.2.5

Data processing

When training the BP neural network, it is necessary to select certain sample data in advance for training, simulation test and verification, and to normalise the data that have been scored with a five-point system to minimise and maximise, so that the data used for network training are within the range of [0, 1]. The equation used is as follows: 23 Xi′=Xi−MIN(X)MAX(X)−MIN(X) where X_i represents a single individual in a sample; MIN(X) and MAX(X) represent the minimum and maximum values in a single individual of the sample, respectively; and Xi′ represents the single individual data of the sample after normalisation.

Then, the classification model of the e-commerce customer based on the BP neural network is constructed as follows (Figure 5):

Fig. 5

The sample data are divided into three categories for training, simulation test and verification of the neural network, while the specific function of the sample is randomly determined by the system. Here, 70% of customer data are taken as the training sample, and the sample size is 182; 15% of customer data are taken as the verification sample, and the sample size is 39; 15% customer data are taken as the test sample, and the sample size is 39. In addition, the number of neurons in the hidden layer is determined; here, it is the default value of 10. Generally, after the default value is selected, the number of neurons can be adjusted according to the actual situation, and different training results can be evaluated, and the number of hidden layers with the smallest error can be selected [32]. Moreover, the number of hidden layers is 6, 7, 8, 9, 10, 11 and 12, respectively. After comparing the mean square error, it is determined that the number of hidden layers is 10. At the same time, according to the network structure diagram, the number of neurons in the input layer is 16 and the number of neurons in the output layer is 1.

Before the neural network training, the maximum training times, the maximum error value and the gradient of the error surface will be set in advance. When one or more indexes reach the set value during the network training, the training will end. Training times is set to the maximum number of iterations, that is, 1000 times, and the final result shows 105 iterations. During the training, the result shows that the training ends in 0.06 seconds, with an initial error of 0.183; at the same time, the initial gradient error of the error surface is 0.368; the maximum number of verification times is 6, and if it exceeds that number, the training will stop, which represents that the training error is greater than the expected error, and the validity verification will stop at once (Table 2).

According to the results, the smaller the value of MSE (performance function), the closer the R value is to 1 and the better the training effect. The results are as follows (Table 3).

Figure 6 shows the error convergence of network training, in which the abscissa represents the training times and the ordinate represents the error value. This training converges at the 105th time, stops training and reaches the minimum error value at the 104th time, which converges to 2.4052 × 10^–6.

Fig. 6

According to the empirical analysis, it can be found that the calculation accuracy of customer value is high. In this study, the trained network has been named net and saved in the workspace of MATLAB software. In practical application, the customer value can be calculated only by using the Subscriber Identity Module (SIM) function [33]. The function of SIM is Y = SIM (X, net), where X represents the numerical value reflected by each index and Y is the calculated customer value. The trained neural network model is used to calculate the customer value that is sorted according to the value, which is divided into four categories, so as to facilitate the subsequent enterprises to formulate targeted customer management measures (Table 4).

The concrete result is that the value of customers in class A is the largest, and the value range is within the range of [0.75, 1]. This kind of customers create the greatest benefits for enterprises, and they will actively purchase from the e-commerce platform and may become loyal customers without excessive publicity and promotion. Customers in class B are of great value, and the value range is within the range of [0.5, 0.75]. This kind of customers can also bring great benefits to enterprises, are interested in e-commerce products and have a higher probability of transforming their value into customers in class A, The value of customers in class C is small, and the numerical range is within the range of [0.25, 0.5]. These customers are sceptical about the way of e-commerce consumption, and they are still in the initial stage of understanding the products of the platform, so the input-to-output ratio of enterprises on this customer will not be very large. Customers in class D have the smallest value, and the value range is within the range of [0, 0.25]. This kind of customers do not trust the consumption pattern of e-commerce, and the benefits they can create for enterprises are very limited. Enterprises are advised to give up on them because they have high requirements on products but show little consumption. Even if enterprises invest high resources in customer development, the customer value they can obtain is still very small.