Explaining Temporal Graph Neural Networks via Feature-induced Information Flow
Ping Xiong, Thomas Schnake, Klaus-Robert Müller, Shinichi Nakajima
Read on arXiv →
Key claim
New method improves interpretability of ETGNNs significantly.
In plain English
Imagine you're trying to understand how a complex system, like a social network or an epidemic model, makes decisions based on events over time. You might use a type of neural network called an event-based temporal graph neural network (ETGNN). These networks are powerful, but they can be really hard to interpret. Current methods for explaining their decisions often miss important connections because they only look at part of the information flow, specifically the event-related embeddings. This can lead to misunderstandings about how different events influence each other and the overall outcome. This is what's called a lack of comprehensive explanation.
To address this, the authors propose a new method that looks at the entire flow of information through all the variables associated with events. This means it considers not just the direct effects of events, but also how they interact with each other over time. By using a framework called the Normalized Relevance Measure (NRM), they can quantify how much each part of the network contributes to the final decision. This method also allows for a deeper analysis of how events relate to one another, which is crucial for understanding long-term dependencies in the data.
In practical terms, this new approach leads to explanations that are not only more accurate but also easier for humans to understand. When tested on synthetic datasets and real-world political event networks, it consistently outperformed existing methods. For anyone building systems that rely on ETGNNs, this means you can now get clearer insights into how your models are making decisions, which is essential for trust and accountability in applications like social network analysis or epidemic tracing.
The proposed method significantly extends existing explanation techniques for complex temporal graph neural networks.
The evaluation on multiple datasets supports the claims, though more extensive baselines could strengthen the findings.
Deep reliability assessment
The methodology supports the claim that tracing event-induced memory/message pathways can recover more faithful explanations on synthetic temporal-graph tasks with known ground-truth chains, and it is plausible on the political-events case qualitatively. The stronger claim that this consistently outperforms existing ETGNN explainers in real deployments is less supported from the provided evidence, especially because the main quantitative ground truth comes from synthetic settings and the real-world evaluation appears harder to verify.
Reproducibility
Partial. The paper says the implementation will be made publicly available upon publication, but no repository URL is provided in the supplied text; synthetic datasets are described procedurally, while full code/data release details are not visible here.
Key figure
Figure 1 contrasts existing ETGNN explainers, which mainly trace relevance from final event-related embeddings to the output, with the proposed Event Relevance framework, which traces event-induced information flow through memory evolution, messages, embeddings, and higher-order event interactions.