A Structural Change Detection Method Integrating Causal Inference and Deep Learning

To accurately detect structural changes driven by shifts in underlying causal mechanisms, we propose CaSCo (Causal Structure Comparison), a novel framework that integrates causal inference with deep representation learning. CaSCo first leverages a differentiable causal structure learning method based on NOTEARS to construct causal graphs from observational data, uncovering the latent causal dependencies among variables. To capture subtle local structural variations and nonlinear patterns, a graph neural network (GNN) using Graph Isomorphism Network (GIN) architecture is employed to encode these causal graphs into expressive embeddings. By comparing the learned causal representations across different time windows or data segments, CaSCo introduces a structural change scoring mechanism that identifies abrupt or gradual changes in the system’s causal dynamics. The framework is general and scalable, making it applicable to tasks such as time-series monitoring, anomaly detection, and system migration analysis. Extensive experiments on both synthetic and real-world datasets demonstrate that CaSCo outperforms existing statistical and deep learning-based approaches in detecting structural changes, particularly under nonstationary environments and high-dimensional conditions.