Concept-Based Interpretability for Spatio-Temporal Graph Neural Network

Modern sensing infrastructures such as industrial IoT systems, energy grids, and cyber-physical networks generate large-scale heterogeneous multivariate time series that are naturally modeled as spatio-temporal graphs. SpatioTemporal Graph Neural Networks (STGNNs) have emerged as a powerful framework for learning over such structured data, achieving strong performance in tasks such as forecasting, imputation, anomaly detection, and classification. However, despite their effectiveness, STGNNs remain largely opaque, limiting their applicability in safety-critical and scientifically sensitive domains where interpretability is essential.

This project is carried out as part of a research visit involving two host institutions: the University of Granada and Imperial College London. It proposes a novel framework for interpretable STGNNs based on concept-based models that explicitly expose semantically meaningful intermediate representations. The goal is to move beyond post-hoc explainability toward inherently interpretable models capable not only of explaining predictions but also of enabling direct intervention and correction. The methodology introduces a concept bottleneck layer tailored to spatio-temporal and graph-structured data, allowing relational and temporal dependencies to be represented in a human-interpretable form.

The research is structured into two phases aligned with the hosting institutions. First, at the University of Granada, under stationary conditions, a concept-based interpretability framework is developed for STGNNs with static graph topology and time-invariant dynamics. Second, at Imperial College London, this framework is extended to non-stationary environments characterized by distribution shifts, concept drift, and evolving graph structures, incorporating physics-informed constraints and drift-aware interpretability mechanisms.

Overall, the project aims to uncover robust spatio-temporal concepts that govern model behavior, enabling transparent, reliable, and controllable STGNNs across both stationary and evolving systems.