An Evaluation Strategy for Commercial Precision Marketing Based on Artificial Neural Network

With the dawn of the Internet era, there is an explosive growth in smart mobile terminals. The China Statistical Report on Internet Development points out that the number of netizens in China surpassed 730 million by the end of 2019, among which 696 million are mobile Internet users. Emerging modern media has overtaken traditional media (e.g. television and newspapers) in the absolute number of viewers and the exposure time [1, 2].

Thanks to the new communication technology of Web 2.0, consumption behaviors have generated a huge amount of data online [3, 4]. Digging deep into the big data on consumption behaviors, enterprises can reasonably predict consumption intentions, pinpoint potential consumers timely, and organize pertinent marketing activities [5-8]. To improve marketing efficiency and conversion rate, enterprises need to effectively explore the consumption demand, and build a complete hierarchical strategy for consumer management [9-12].

Precision marketing refers to the marketing targeted at consumers with strong willingness to consume. To realize precision marketing, it is necessary to create a complete consumer database, identify the enthusiastic consumers, and implement consumer-centered marketing activities [13-15]. Lin et al. [16] mentioned that, during precision marketing, enterprises could establish a precise and controllable mechanism for marketing and communication based on consumption actions and results. Kanmani and Jayapradha [17] considered precise consumer positioning the key to maintain long-term consumer relationships, suggested improving consumer loyalty and stickiness by preparing targeted personalized plans, and advised enterprises to maximize consumer value through the provision of enriched and refined services. Alkhatib and Abualigah [18] constructed consumer migration and retention models to measure consumer value by the net cash flow. After proposing a new concept called consumer lifetime value, Xiong and Gao [19] expounded the key components of consumer value, classified consumers by purchase behaviors and statistical features, and put forward targeted marketing plans.

The rapid development of computer technology has promoted the application of artificial neural network (ANN) in various industries [20, 21]. Anitha and Patil [22] combined the Logit model and neural network to classify consumers, and achieved a classification accuracy of 86.5%. To evaluate consumer credit, Monalisa et al. [23] established a backpropagation neural network (BPNN), and demonstrated the high accuracy of the network in handling massive data and predicting potential purchase intentions of consumers. Haghighatnia et al. [5] applied the optimized fuzzy neural network (FNN) to evaluate e-commerce consumers and predict their purchase intentions, and proved that the optimized FNN is a scientific and accurate network with good nonlinearity, self-learning ability, self-organization.

So far, mature theories have been proposed on consumer value and precision marketing. But further research is needed to mine the consumer demand and its value generated in commercial precision marketing. The ANN has achieved good effects in data classification and screening. However, there is little report that evaluates the effect of precision marketing based on evaluation index system (EIS). Therefore, this paper puts forward an evaluation strategy for commercial precision marketing based on the ANN.

The remainder of this paper is organized as follows: Section 2 presents an EIS for commercial precision marketing based on improved attention-interest-desire-memory-action (ADIMA) model, and determines the principal evaluation indices through principal component analysis (PCA); Section 3 establishes an ANN evaluation model for commercial precision marketing, optimizes the model through k-means clustering (KMC), and realizes the optimized model on MATLAB; Section 4 verifies the reasonability of the proposed EIS and the effectiveness of the proposed model; Section 5 puts forward the conclusions.

3.1 Model construction

Considering their numerous attributes and correlations, the evaluation indices for commercial precision marketing cannot be directly used in cluster analysis. Otherwise, the information will overflow, and undermine the clustering effect. To solve the problem, this paper builds up an ANN evaluation model for commercial precision marketing. The ANN was adopted for the cluster analysis of the principal indices identified in PCA, providing a reference for decision-makers on commercial marketing and consumer relationship management. Figure 2 explains the workflow of model construction and application. The structure of the model is illustrated in Figure 3. It can be seen that the model is a BPNN, consisting of an input layer, a hidden layer, and an output layer.

2.png

Figure 2. The construction and application of the ANN model

3.png

Figure 3. The structure of the ANN model

The composite score of commercial precision marketing was imported to the ANN. Let j be the serial number of index sample, B_ij be the input matrix on the input layer, and γ_is be the connection weight between the i-th input layer node and the s-th hidden layer node. Then, the activation function of the hidden layer can be expressed as:

$h_{s j}=\frac{1}{1+\left[\left(\sum_{i=1}^{m} \gamma_{i s} B_{i j}\right)^{-1}-1\right]^{2}}$     (20)

where, h_sj is the input matrix of the hidden layer; η_sv is the connection weight between the s-th hidden layer node and the v-th output layer node; y_vj is the output of the v-th output layer node:

$y_{v j}=\frac{1}{1+\left[\left(\sum_{s=1}^{P} \eta_{s v} h_{s j}\right)^{-1}-1\right]^{2}}$     (21)

3.2 Model optimization

In this sub-section, the BPNN is optimized through KMC. In addition, the number and width of hidden layer nodes was determined, as well as the position of cluster heads. Furthermore, the output weights of the BPNN were adjusted by gradient descent. The entire optimization process is shown in Figure 4.

4.png

Figure 4. The workflow of KMC optimization of BPNN

Step 1. Normalize the sample data by formulas (10)-(12). Then, compute the densities of N data samples on the same layer by:

$C_{1}=\sum_{i=1}^{n} \exp \left[-\frac{\left\|B_{i j}-B_{i k}\right\|^{2}}{r^{2}}\right]$     (22)

where, r is a constant about the radius of the neighborhood of index B_ij. Then, choose the sample with the highest density as the first cluster head C₁.

Further solve and revise the densities of the remaining N-1 samples by:

$C_{u}=C_{u}-C_{u-1} \exp \left[-\frac{\left\|B_{i j}-B_{u-1}\right\|^{2}}{r^{2}}\right]$     (23)

where, Ḃ_u-1 is the sample data corresponding to the u-1-th cluster head. Then, choose the sample with the highest density as the u-th cluster head C_u.

Step 2. Keep repeating Step 1 until finding K cluster heads C₁, C₂, …, C_K. Compute the distance from each index sample to each cluster head by:

$d_{u}=\left\|B_{i j}-C_{u}\right\|$     (24)

Step 3. Classify the index samples by:

$\Gamma\left(B_{i j}\right)=\min d_{u}=\min \left\|B_{i j}-C_{u}\right\|$     (25)

Optimize and finetune the cluster head of each class by:

$C_{u}^{\prime}=\frac{1}{N} \sum_{j=1}^{N_{t}} B_{i j}$     (26)

where, N_i is the number of index samples of class t. If the finetuned cluster heads are not consistent with the original ones, re-calculate the distance from each index sample to each cluster head; otherwise, go to Step 4.

Step 4. Compute the width of each hidden layer node by:

$\varphi_{u}=\varepsilon \min d_{u}$     (27)

where, ε=1 is the overlap coefficient. The width of a hidden layer node is the shortest distance between the u-th cluster head and other cluster heads.

Step 5. Calculate the deviation of expected output from the actual output:

error $=\sum_{i=1}^{N}$ error $_{i}=\frac{1}{2} \sum_{i=1}^{N}\left[y^{*}\left(B_{i j}\right)-y\left(B_{i j}\right)\right]^{2}$     (28)

where, y^*(B_ij) and y(B_ij) are the expected output and actual output of the BPNN based on the input of B_ij.

To minimize the deviation, revise the connection weight matrix between hidden layer and output layer by gradient descent:

$\eta_{s v}^{\prime}=\eta_{s v}-\delta\left(\frac{1}{2} \sum_{s=1}^{P} \frac{\partial E}{\partial \eta_{s v}}\right)$     (29)

where, δ=0.001 is the learning rate.

3.3 Model implementation

The BPNN evaluation model for commercial precision marketing mainly involves three parts: index samples, data normalization algorithm, and online model. Specifically, the index database for commercial precision marketing was connected with the database in MATLAB, so that the data can be directly read. The connection command can be expressed as:

Conn=database('','root','','com.mysql.jdbc.Driver','jdbc:mysql://localhost/db_name?characterEncoding=Data In-10')

The statement of the samples of each index can be read by:

t=exec(conn,’select*from data’);

The data normalization algorithm was realized by calling the premnmx function of MATLAB:

$\left[B_{i j}^{\prime}, B_{i j-\min }^{\prime}, B_{j-\max }^{\prime}, B_{i j}^{\prime \prime}, B_{i j-\min }^{\prime \prime}, B_{j-\max }^{\prime \prime}\right]$$=\operatorname{premnmx}\left(B_{i j}^{\prime}, B_{i j}^{\prime \prime}\right)$     (30)

The ANN algorithm was realized by calling the nwerb function of MATLAB:

$[\mathrm{Net}, \mathrm{Tr}]=\operatorname{newr} b\left(B_{i j}, y^{*}\left(B_{i j}\right), M S E, E S, P_{\max }\right)$     (31)

where, MSE is the preset mean squared error of training; ES is the expansion speed of the ANN; P_max is the maximum number of input layer nodes; Net is the established BPNN; T_r is the training record of the BPNN.

The KMC algorithm was realized by calling the kmeans function in MATLAB:

$\left[I d x, N_{t}, C\right]=k \operatorname{means}\left(B_{i j}, K, I C\right)$     (32)

where, IC is the initial cluster head; Idx is the class of an index sample; N_t is the number of index samples in each class.