GINTRIP: Interpretable Temporal Graph Regression using Information bottleneck and Prototype-based method

TL;DR

GINTRIP tackles interpretability in temporal graph regression by marrying the Information Bottleneck framework with prototype-based learning. The method derives tractable MI bounds and introduces learnable prototypes to obtain time-aware subgraphs that support both forecasting and explanation, aided by an unsupervised auxiliary classification head for diverse concepts. Empirical results on traffic and crime datasets show that GINTRIP achieves superior forecasting accuracy and higher interpretability fidelity compared to state-of-the-art baselines, while visualizations confirm coherent, time-resolved subgraphs and prototype semantics. This work advances self-explainable spatio-temporal GNNs with practical impact for transportation and public safety analytics, and it opens avenues for dynamic and hypergraph interpretability in future research.

Abstract

Deep neural networks (DNNs) have demonstrated remarkable performance across various domains, but their inherent complexity makes them challenging to interpret. This is especially true for temporal graph regression tasks due to the complex underlying spatio-temporal patterns in the graph. While interpretability concerns in Graph Neural Networks (GNNs) mirror those of DNNs, no notable work has addressed the interpretability of temporal GNNs to the best of our knowledge. Innovative methods, such as prototypes, aim to make DNN models more interpretable. However, a combined approach based on prototype-based methods and Information Bottleneck (IB) principles has not yet been developed for temporal GNNs. Our research introduces a novel approach that uniquely integrates these techniques to enhance the interpretability of temporal graph regression models. The key contributions of our work are threefold: We introduce the Graph INterpretability in Temporal Regression task using Information bottleneck and Prototype (GINTRIP) framework, the first combined application of IB and prototype-based methods for interpretable temporal graph tasks. We derive a novel theoretical bound on mutual information (MI), extending the applicability of IB principles to graph regression tasks. We incorporate an unsupervised auxiliary classification head, fostering diverse concept representation using multi-task learning, which enhances the model's interpretability. Our model is evaluated on real-world datasets like traffic and crime, outperforming existing methods in both forecasting accuracy and interpretability-related metrics such as MAE, RMSE, MAPE, and fidelity.

Figures (7)

• Figure 1: Proposed method architecture. It includes a $GNN$ encoder, subgraph extractor, learnable prototypes, regression, and classification head components. Overall, the temporal graph passes to the $GNN$ model, and the latent representation {$h_{1},h_{2},...,h_{N}$} is processed by $G_{sub}$$Extractor$ to extract a subgraph $G_{sub}$. A pooling block is used to capture global-level features. A similarity function $\gamma$ computes similarities between learnable prototypes ${G}_{p} = \left\{ \mathbf{z}_{G_{p}}^1, \mathbf{z}_{G_{p}}^2, \cdots, \mathbf{z}_{G_{p}}^M \right\}$ and subgraph representation $Z_{sub}$. The concatenated similarities {${S^1,S^2,...,S^M}$} with $Z_{sub}$ are used for the proposed pseudo-label classification and the forecasting task.
• Figure 2: Positive fidelity over k. High fidelity is an indicator of interpretability.
• Figure 3: Positive over the negative fidelity based on the two selected subgraphs regarding the k.
• Figure 4: Explanation of the learned prototypes for a spatio-temporal prediction example. The y-axis denotes traffic speed.
• Figure 5: Crime rate prediction benchmarks for Chicago and New York City.