Fine-grained Graph Rationalization

TL;DR

This work addresses the challenge of extracting meaningful graph rationales by moving from coarse graph-level intervention to fine-grained, node-level and virtual node-level interventions. It introduces FIG, a Transformer-inspired framework with four modules (encoder, augmenter, intervener, predictor) and two variants (FIG-N, FIG-VN) that enable adversarial, min–max training to ensure rationale invariance under environment changes. Empirical results on 7 real-world datasets against 13 baselines show significant performance gains and robust efficiency, validating the value of fine-grained rationalization for graph property prediction and explainability. The approach advances graph rationale discovery by leveraging attention-based interactions at finer granularity, with practical implications for reliable, interpretable graph models in diverse domains.

Abstract

Rationale discovery is defined as finding a subset of the input data that maximally supports the prediction of downstream tasks. In the context of graph machine learning, graph rationale is defined to locate the critical subgraph in the given graph topology. In contrast to the rationale subgraph, the remaining subgraph is named the environment subgraph. Graph rationalization can enhance the model performance as the mapping between the graph rationale and prediction label is viewed as invariant, by assumption. To ensure the discriminative power of the extracted rationale subgraphs, a key technique named "intervention" is applied whose heart is that given changing environment subgraphs, the semantics from the rationale subgraph is invariant, guaranteeing the correct prediction result. However, most, if not all, of the existing graph rationalization methods develop their intervention strategies on the graph level, which is coarse-grained. In this paper, we propose fine-grained graph rationalization (FIG). Our idea is driven by the self-attention mechanism, which provides rich interactions between input nodes. Based on that, FIG can achieve node-level and virtual node-level intervention. Our experiments involve 7 real-world datasets, and the proposed FIG shows significant performance advantages compared to 13 baseline methods.

Figures (7)

• Figure 1: Illustration of the rationale/environment decomposition and intervention. Round nodes denote graph rationales, and square nodes (with stripes) denote the environments. The intervention aims to ensure the rationale from graph $\mathcal{G}$ truly has the discriminative power for the label $y_{\mathcal{G}}$.
• Figure 2: Pipeline comparison between existing work GREA and proposed FIG. $\circ$ denotes function composition. GREA designs the intervention at the graph level, and the proposed FIG designs the intervention at the node/virtual node level. The augmented environment $\tilde{\mathbf{H}}_{\texttt{env}}$ is from another graph $\tilde{\mathcal{G}}$ (through the Encoder and Augmenter) in the batch.
• Figure 3: FIG-N single training step for every training graph $\mathcal{G}$
• Figure 4: Performance of FIG-N/VN with different $\hat{K}$.
• Figure 5: Training loss of FIG-N/VN with different datasets and encoders.