STX-Search: Explanation Search for Continuous Dynamic Spatio-Temporal Models

TL;DR

STX-Search addresses the challenge of explaining predictions from continuous dynamic spatio-temporal models by formulating a subset selection problem on the computation graph $\mathcal{G}^{t_k}_c$ and solving it with simulated annealing. It introduces a novel multi-stage objective that jointly optimizes fidelity to the full model, a new $\alpha\text{Fidelity}$ measure, and sparsity to enhance interpretability. The method applies to static and dynamic graphs and is evaluated against TGNNExplainer and Temp-ME on Wikipedia and Reddit, showing superior MAE and $\alpha\text{Fidelity}$ and enabling automatic determination of explanation size. This work advances trustworthy, interpretable spatio-temporal forecasting by providing compact, faithful explanations for complex dynamic graphs.

Abstract

Recent improvements in the expressive power of spatio-temporal models have led to performance gains in many real-world applications, such as traffic forecasting and social network modelling. However, understanding the predictions from a model is crucial to ensure reliability and trustworthiness, particularly for high-risk applications, such as healthcare and transport. Few existing methods are able to generate explanations for models trained on continuous-time dynamic graph data and, of these, the computational complexity and lack of suitable explanation objectives pose challenges. In this paper, we propose $\textbf{S}$patio-$\textbf{T}$emporal E$\textbf{X}$planation $\textbf{Search}$ (STX-Search), a novel method for generating instance-level explanations that is applicable to static and dynamic temporal graph structures. We introduce a novel search strategy and objective function, to find explanations that are highly faithful and interpretable. When compared with existing methods, STX-Search produces explanations of higher fidelity whilst optimising explanation size to maintain interpretability.

Figures (4)

• Figure 1: An illustration showcasing the dependency between 2 users within a social network setting. The information regarding the connection between users 3 and 4 may be crucial in predicting the state of the connection between users 1 and 2.
• Figure 2: Two examples of an explanation, $\mathcal{R}^{t_k}$, and its complement, $\mathcal{G}^{t_k}_c \backslash \mathcal{R}^{t_k}$, which when combined give the computation graph. Both explanations contain a number of important events. The explanation on the left contains two groups of dependent events, with one of them highlighted in blue. The explanation on the right no longer contains one of the dependent events in the highlighted pairing, which is now part of the explanation complement, $\mathcal{G}^{t_k}_c \backslash \mathcal{R}^{t_k}$. The explanation with the dependent events grouped together is more faithful to the base model prediction, and has a lower Fidelity$^-$. However, the less faithful explanation on the right incorrectly results in a higher $\Delta$Fidelity.
• Figure 3: A comparison of the average MAE and $\alpha\text{Fidelity}$ achieved by each method when generating explanations of different sizes for 100 random instances from each dataset.
• Figure 4: A comparison of average explanation MAE and $\alpha\text{Fidelity}$ achieved by STX-Search when performing a multi-stage search to automatically find the best explanation size for 100 random instances from the Wikipedia dataset to explain a TGN base model using different $\lambda$ values in the search objective function. The distribution of MAE and $\alpha\text{Fidelity}$ against explanation size is also shown.

Theorems & Definitions (1)

• Definition 2.1