Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-Explainable Temporal Graph Networks based on Graph Information Bottleneck

Sangwoo Seo, Sungwon Kim, Jihyeong Jung, Yoonho Lee, Chanyoung Park

TL;DR

The paper tackles the challenge of explainability in temporal graphs, where existing post-hoc explanations are inefficient and brittle to model updates. It introduces TGIB, a Self-Explainable Temporal Graph Network that uses Graph Information Bottleneck to extract a time-aware bottleneck subgraph $\mathcal{R}^k$ from the past, explaining the prediction of the target event $e_k$ while predicting its occurrence. Core contributions include (i) a GIB-based objective combining $\mathcal{L}_{cls}$ and $\mathcal{L}_{MI}$ for end-to-end learning, (ii) time-aware event representations with explicit temporal encoding, (iii) a tractable MI approximation via Bernoulli sampling and Gumbel-Softmax, and (iv) a spurious-correlation removal mechanism to improve generalization. Empirically, TGIB achieves state-of-the-art link prediction performance and superior explanation quality across six real-world temporal graphs, while offering efficient, integrated training without the need for retraining a separate explanation model.

Abstract

Temporal Graph Neural Networks (TGNN) have the ability to capture both the graph topology and dynamic dependencies of interactions within a graph over time. There has been a growing need to explain the predictions of TGNN models due to the difficulty in identifying how past events influence their predictions. Since the explanation model for a static graph cannot be readily applied to temporal graphs due to its inability to capture temporal dependencies, recent studies proposed explanation models for temporal graphs. However, existing explanation models for temporal graphs rely on post-hoc explanations, requiring separate models for prediction and explanation, which is limited in two aspects: efficiency and accuracy of explanation. In this work, we propose a novel built-in explanation framework for temporal graphs, called Self-Explainable Temporal Graph Networks based on Graph Information Bottleneck (TGIB). TGIB provides explanations for event occurrences by introducing stochasticity in each temporal event based on the Information Bottleneck theory. Experimental results demonstrate the superiority of TGIB in terms of both the link prediction performance and explainability compared to state-of-the-art methods. This is the first work that simultaneously performs prediction and explanation for temporal graphs in an end-to-end manner.

Self-Explainable Temporal Graph Networks based on Graph Information Bottleneck

TL;DR

The paper tackles the challenge of explainability in temporal graphs, where existing post-hoc explanations are inefficient and brittle to model updates. It introduces TGIB, a Self-Explainable Temporal Graph Network that uses Graph Information Bottleneck to extract a time-aware bottleneck subgraph from the past, explaining the prediction of the target event while predicting its occurrence. Core contributions include (i) a GIB-based objective combining and for end-to-end learning, (ii) time-aware event representations with explicit temporal encoding, (iii) a tractable MI approximation via Bernoulli sampling and Gumbel-Softmax, and (iv) a spurious-correlation removal mechanism to improve generalization. Empirically, TGIB achieves state-of-the-art link prediction performance and superior explanation quality across six real-world temporal graphs, while offering efficient, integrated training without the need for retraining a separate explanation model.

Abstract

Temporal Graph Neural Networks (TGNN) have the ability to capture both the graph topology and dynamic dependencies of interactions within a graph over time. There has been a growing need to explain the predictions of TGNN models due to the difficulty in identifying how past events influence their predictions. Since the explanation model for a static graph cannot be readily applied to temporal graphs due to its inability to capture temporal dependencies, recent studies proposed explanation models for temporal graphs. However, existing explanation models for temporal graphs rely on post-hoc explanations, requiring separate models for prediction and explanation, which is limited in two aspects: efficiency and accuracy of explanation. In this work, we propose a novel built-in explanation framework for temporal graphs, called Self-Explainable Temporal Graph Networks based on Graph Information Bottleneck (TGIB). TGIB provides explanations for event occurrences by introducing stochasticity in each temporal event based on the Information Bottleneck theory. Experimental results demonstrate the superiority of TGIB in terms of both the link prediction performance and explainability compared to state-of-the-art methods. This is the first work that simultaneously performs prediction and explanation for temporal graphs in an end-to-end manner.
Paper Structure (28 sections, 1 theorem, 28 equations, 4 figures, 8 tables, 1 algorithm)

This paper contains 28 sections, 1 theorem, 28 equations, 4 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Temporal Graph Neural Networks
  4. Graph Information Bottleneck
  5. GNN Explainability
  6. Preliminaries
  7. Temporal Graph Model
  8. Graph Information Bottleneck
  9. Methodology
  10. GIB-based Objective for Temporal Graph
  11. Time-aware event representation
  12. Minimizing $I(\mathcal{R}^k ; e_k, \mathcal{G}^k)$ in Equation \ref{['eq:compression']}
  13. Minimizing $-I(Y_k ; \mathcal{R}^k)$ in Equation \ref{['eq:prediction']}
  14. Spurious Correlation Removal
  15. Experiments
  16. ...and 13 more sections

Key Result

Proposition 1

Assume that each $\mathcal{G}^k$ contains $\mathcal{R}^{k*}$, which determines $Y_k$. In other words, for some deterministic invertable function $f$ with randomness $\epsilon$ that is independent of $\mathcal{G}^k$, it satisfies $Y = f(\mathcal{R}^{k*}) + \epsilon$. Then, for any $\beta \in [0,1]$,

Figures (4)

  • Figure 1: Comparison between T-GNNexplainer and TGIB.
  • Figure 2: The architecture of our proposed TGIB.
  • Figure 3: Spurious Correlation Removal of TGIB.
  • Figure 4: Comparison of explanation visualization for explanation models for static graphs and TGIB.

Theorems & Definitions (1)

  • Proposition 1