[3O1-GS-10v] AI応用：実世界システム

10:00 〜 10:15

[3O1-GS-10v-05] Customer Churn Prediction via Integration of TDA and GNN: Evaluation and Interpretation of Structural Robustness based on Persistent Homology

〇Yifan Song¹, Tengfei Shao¹, Masayuki Goto¹ (1. Waseda University)

キーワード：

Customer Churn Prediction、Graph Neural Networks、Topological Data Analysis、Persistent Homology、Structural Robustness

In the subscription economy, where customer retention drives enterprise value, Graph Neural Networks (GNNs) have emerged as powerful tools for churn prediction. However, most existing GNN methods rely on feature "smoothing" and aggregating local adjacencies, often missing higher-order topological features such as the holistic "shape" and "structural robustness" of user networks. This study proposes a novel framework integrating Topological Data Analysis (TDA) with GNNs to quantify local structural robustness, enhancing predictive performance and interpretability. We employ a Relative Neighborhood Graph (RNG) to prune noisy connections and build an essential user network, then apply Persistent Homology (PH) to each user's neighborhood to extract the "number of connected components," a metric of structural fragmentation, and integrate it into the GNN. Experiments show the method outperforms baselines. Crucially, high-risk users tend to have more connected components, indicating their local networks lack cohesiveness and are fragmented into scattered clusters. These findings suggest incorporating topological features improves accuracy and enables structural interpretation of churn: "fragmentation leads to churn," offering valuable insights for retention strategies.