Customer Churn Prediction Based on Multiscale Attention and Convolutional Neural Networks

Most of the feature extraction methods for telecommunication customer churn research tend to lose important information from other scales by stacking the single-scale receptive field information, which leads to limited prediction effect. To address this problem, this paper proposes a customer churn risk prediction model based on a multi-scale convolutional attention module-convolutional neural network. Firstly, the features are reshaped into a form similar to grey scale images, thus allowing the 2D convolutional neural network to extract local spatial features, and using residual blocks to enhance the information flow of the deep network. Based on this, a multi-scale convolutional attention module is designed in this paper to effectively capture multi-scale feature information and enhance feature diversity from different domains, while adaptively learning the weights of each feature channel and feature map pixel point. Ten-fold cross-validation experiments are conducted on two publicly available benchmark datasets, and in order to enhance the recognition ability of the proposed model for minority class instances, SMOTENN is introduced to over-and under-sample the telecom datasets used, and the results show that the proposed model outperforms the comparison model in terms of the scores on Accuracy, Precision, Recall, and F1-score, providing a better decision-making reference for telecom service providers to identify churn.