Research on Interpretable Customer Churn Prediction Based on Attention Mechanism

Customer churn prediction is critical to an enterprise. Therefore, the improvement of the churn prediction model can directly help enterprises to better manage customers to obtain more profit. Meanwhile, the explanation of the decision-making mechanism of the churn prediction model can guide enterprises to accurately understand and evaluate the likelihood of customer churn, and then make more targeted measures to prevent customer churn. However, most existing deep learning approaches have poor model interpretability due to the black-box characteristics of neural networks. In addition, there are few previous studies on the model interpretability of customer churn predictions. Accordingly, in this paper, we propose a novel model named Interpretable DeepFM (iDeepFM) to study the model interpretability. The Proposed iDeepFM model introduces multi-head self-attention mechanism to the embedding layer, the linear interaction component, and the deep component. Especially, the deep component computes the high-order feature interactions by stacking multiple attention blocks. Furthermore, we conduct extensive experiments using two real-world telecom customer churn datasets. We calculate attention scores which are used to interpret the prediction results for each component. The results show that the proposed approach not only improves the performance of DeepFM but also offers good model interpretability.