Explaining Temporal Graph Neural Networks via Feature-induced Information Flow

2026-06-25 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new attribution method is proposed to explain Event-based Temporal Graph Neural Networks (ETGNNs), which are widely used in social network analysis, epidemic tracing, recommender systems, and political event forecasting. Current explanation techniques for ETGNNs are limited, focusing only on event-related embeddings and missing crucial information flow through event-induced variables that mediate node interactions and capture long-range temporal dependencies. The novel method analyzes the entire information flow via all event-associated variables. It extends the Normalized Relevance Measure (NRM) framework with a modular decomposition procedure to handle complex neural architectures. This allows explicit quantification of information flow from both event embeddings and event-induced variables, ensures latent variable comparability across layers, and supports higher-order interaction analysis. Evaluations on two synthetic datasets (epidemic tracing, social dynamics) and a real-world political event network dataset demonstrate its superior performance and human-interpretable explanations compared to existing approaches.

Key takeaway

For Machine Learning Engineers developing or deploying Event-based Temporal Graph Neural Networks, this new attribution method offers significantly improved explainability. You should consider integrating this NRM-based approach to gain deeper insights into model decisions. This is crucial for understanding long-range temporal dependencies and complex node interactions. It enhances trustworthiness in sensitive applications like epidemic tracing or political forecasting.

Key insights

The new method explains ETGNNs by analyzing the entire information flow, including event-induced variables, for more interpretable results.

Principles

• Event-induced variables are critical for ETGNN explainability.
• Entire information flow analysis improves explanation quality.
• Modular decomposition aids complex neural architecture analysis.

Method

The method extends the Normalized Relevance Measure (NRM) framework with modular decomposition to quantify information flow from event embeddings and through event-induced variables, ensuring latent variable comparability and supporting higher-order interaction analysis.

In practice

• Apply to social network analysis.
• Use for epidemic tracing models.
• Enhance recommender system transparency.

Topics

• Temporal Graph Neural Networks
• Explainable AI
• Attribution Methods
• Normalized Relevance Measure
• Information Flow Analysis

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Temporal Functional Circuits: From Spline Plots to Faithful Explanations in KAN Forecasting — Temporal Functional Circuits enable mechanistic explanations in KAN-based time-series forecasting by transforming latent edge functions into explicit, temporally grounded insights.
• Towards Rigorous Explainability by Feature Attribution — Symbolic XAI methods offer rigorous alternatives to non-symbolic approaches for reliable feature importance attribution.
• Forecasting Is Not Attribution: Localizing Decoder Bypass in Graph-Based Neural Marketing Mix Models — Forecasting accuracy in neural MMMs does not guarantee correct attribution, often hiding a "decoder bypass" failure.
• Applied Explainability for Large Language Models: A Comparative Study — Gradient-based explainability provides stable LLM insights, while attention is efficient but less aligned with predictions.
• Beyond Accuracy: Introducing a Symbolic-Mechanistic Approach to Interpretable Evaluation — Mechanism-aware evaluation using symbolic rules and interpretability better distinguishes true generalization from pattern exploitation.
• Graph networks Frozen in Time Are Useless. Temporal Graph Networks Aren’t. — Temporal Graph Networks (TGNs) enable dynamic graph analysis by giving nodes memory and encoding temporal interactions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.