Application of backpropagation neural network algorithm in e-commerce customer churn prediction

Abstract

In the increasingly dominant consumer market of contemporary e-commerce, it has become urgent for enterprises to have a deeper and more comprehensive understanding of their e-commerce customers. A new method combining the K-means clustering algorithm with a backpropagation neural network has been proposed in the study. Specifically, customer data are first preprocessed through K-means clustering. Then, the initial weight distribution of the data is determined based on the clustering results to optimize the initial weight setting of the backpropagation neural network, thereby accelerating convergence during model training and effectively avoiding the problem of getting stuck in local optima. The results showed that the use of K-means clustering on customer information had a greater impact on the prediction ability of the algorithm model. The Geometric Means (G-means) and F1-values of the prediction model combining K-means clustering and backpropagation neural network were about 15 % and 31 % higher than those of the C4.5 decision tree algorithm, Support Vector Machine (SVM), and Logistic Regression algorithms, respectively. The area under the curve, accuracy, precision, and recall were 0.045, 0.037, 0.042, and 0.021 higher than those of backpropagation neural networks, respectively. The experimental results meet expectations, indicating that the proposed model is suitable for predicting customer churn in e-commerce.