Customer Churn Prediction Based on Coordinate Attention Mechanism with CNN-BiLSTM

Due to increased competition in the marketplace, companies in all industries are facing the problem of customer attrition. In order to expand their market share and increase profits, companies have shifted from the concept of ‘acquiring new customers’ to ‘retaining old customers’. In this study, we design a deep learning model based on multi-network feature extraction and an attention mechanism, convolutional neural network–bidirectional long and short-term memory network–fully connected layer–coordinate attention (CNN-BiLSTM-FC-CoAttention), and apply it to customer churn risk assessment. In the data preprocessing stage, the imbalanced dataset was processed using the SMOTE-ENN hybrid sampling method. In the feature extraction stage, a sequence-based CNN and time-based BiLSTM are combined to extract the local and time series features of the customer data. In the feature transformation stage, high-level features are extracted using a fully connected layer of 64 Relu neurons and the sequence features are reshaped into matrix features. In the attention enhancement stage, the extracted feature information is refined using a coordinate attention learning module to fully learn the channel and spatial location information of the feature map. To evaluate the performance of the proposed model, we include public datasets from telecom, bank and insurance industries for ten-fold cross-validation experiments, and the results show that the CNN-BiLSTM-FC-CoAttention model outperforms the comparison models in all metrics. Our proposed model improves the accuracy and generalisation of the model prediction by combining multiple algorithms, enabling it to be widely used in multiple industries. As a result, the model gives enterprises a better and more general decision-making reference for the timely identification of potential churn customers.