Causality-Aware Recursive Encoding for Interpretable Temporal Knowledge Graph Extrapolation

Predicting future anomalies in Industrial Control System (ICS) sensor networks is critical for preventing equipment failures, minimizing downtime, and ensuring operational safety. Temporal Knowledge Graphs (TKGs) offer a structured framework to model evolving causal relationships among sensors, actuators, and processes within these infrastructures. However, state-of-the-art subgraph-based TKG extrapolation methods are susceptible to spurious correlations introduced by noisy data and lack the ability to capture essential multi-hop relational semantics required to model complex fault propagation chains in ICS networks. To address this limitation, we propose CARE (Causality-Aware Recursive Encoding), a framework for robust TKG reasoning in ICS settings. CARE integrates causal subgraphs to isolate structurally essential dependencies for sensor anomaly prediction, while filtering non-causal features such as temporary sensor correlations without genuine links. A subgraph-aware recursive encoder, inspired by path aggregation algorithms, captures multi-hop path semantics through transitively aware message passing. This design enables the modeling of long-range dependencies in fault propagation, where primary sensor anomalies propagate through interconnected control loops. Applied to ICS sensor anomaly prediction, CARE outperforms state-of-the-art methods in evaluations on real-world sensor data. Moreover, by providing traceable causal paths, CARE supports interpretability for proactive root cause diagnosis, achieving a 3.44% MRR improvement on ICEWS18 and superior performance on ICS-specific datasets.