Explanation-Preserving Augmentation for Semi-Supervised Graph Representation Learning

Conference: AAAI 2026 arXiv: 2410.12657 Code: https://github.com/realMoana/EPA-GRL Area: Self-Supervised Learning Keywords: Graph Contrastive Learning, Semantics-Preserving Augmentation, Explainable AI, Semi-Supervised Learning, GNN Explainer

TL;DR

This paper proposes EPA-GRL (Explanation-Preserving Augmentation for Graph Representation Learning), which employs a GNN explainer trained with a small number of labels to identify semantic subgraphs (explanation subgraphs). During augmentation, only the non-semantic portions (marginal subgraphs) are perturbed, achieving semantics-preserving graph augmentation. EPA-GRL significantly outperforms semantics-agnostic random augmentation methods across 6 benchmarks.

Background & Motivation

Background: Graph Contrastive Learning (GCL) learns invariant representations by generating two augmented views of the same graph. Methods such as GraphCL and JOAO adopt augmentation strategies including random node/edge dropping.

Limitations of Prior Work: Existing augmentation strategies are semantics-agnostic — random perturbations may destroy key substructures relevant to classification (e.g., benzene rings in molecular graphs), causing augmented graphs to lose core semantics and degrading downstream classification performance.

Key Challenge: Effective augmentation must simultaneously satisfy (1) semantic preservation and (2) introduction of variance; however, existing methods focus exclusively on the latter.

Goal: How can graph augmentation preserve semantics while introducing sufficient variance?

Key Insight: Leverage GNN explainability techniques to identify semantic subgraphs, protecting them during augmentation while perturbing only the non-semantic parts.

Core Idea: Train an explainer with a small number of labels → identify semantic subgraphs → perturb only the non-semantic portions = semantics-preserving graph augmentation.

Method

Overall Architecture

EPA-GRL is a two-stage approach: 1. Pre-training Stage: Train a GNN classifier \(f_{pt}\) and explainer \(\Psi\) using a small number of labeled graphs. 2. Representation Learning Stage: Use the frozen explainer to generate semantic subgraphs → perturb only the marginal part → train the encoder via contrastive learning.

Key Designs

1. GNN Explainer Pre-training:

   • Function: Learn a parametric explainer that takes a graph \(G\) as input and outputs an explanation subgraph \(G^{(exp)}\).
   • Mechanism: Train under the Graph Information Bottleneck (GIB) principle: \(\arg\min_{\Psi} \sum CE(Y; f_{pt}(\Psi(G))) + \lambda|\Psi(G)|\). The first term ensures that the subgraph retains classification-relevant information; the second enforces compactness (preventing inclusion of irrelevant components).
   • Design Motivation: With a small number of labels, the explainer learns to identify class-discriminative substructures and generalizes to unlabeled graphs.

2. Explanation-Preserving Augmentation:

   • Function: Generate augmented graphs that preserve semantics.
   • Mechanism: \(G^{(exp)} = \Psi(G)\) (kept unchanged); \(\Delta G = G \setminus G^{(exp)}\) (randomly perturbed). Perturbation strategies include: node dropping on \(\Delta G\), edge dropping on \(\Delta G\), attribute masking on \(\Delta G\), subgraph sampling from \(\Delta G\), and mixup (replacing the current marginal subgraph with that of another graph).
   • Design Motivation: The semantic subgraph is protected from corruption, while the marginal part provides sufficient variance.

3. Theoretical Analysis:

   • Function: Formally establish the superiority of semantics-preserving augmentation.
   • Mechanism: On a modified BA-2motifs graph, it is proven that a semantics-preserving encoder \(f_{enc}^{sp}\) achieves classification error approaching 0, whereas a semantics-agnostic encoder \(f_{enc}^{sa}\) achieves error approaching 1/2 (random guessing) — the gap can be made arbitrarily large.

Loss & Training

• The explainer is trained with GIB loss combined with cross-entropy (CE) loss.
• Representation learning employs GraphCL (contrastive) or SimSiam loss.
• Semi-supervised setting: a small number of labels are used to train the explainer, while all graphs (including unlabeled ones) participate in contrastive learning.

Key Experimental Results

Main Results

Graph classification accuracy on 6 benchmarks (BA-2motifs, MUTAG, PROTEINS, DD, NCI1, COLLAB):

Method	BA-2motifs	MUTAG	PROTEINS	NCI1
GraphCL (random aug)	Baseline	Baseline	Baseline	Baseline
JOAO	Marginal gain	Marginal gain	Marginal gain	Marginal gain
AD-GCL	Moderate	Moderate	Moderate	Moderate
EPA-GraphCL	Significant gain	Significant gain	Significant gain	Significant gain
EPA-SimSiam	Significant gain	Significant gain	Significant gain	Significant gain

Ablation Study

Configuration	Performance	Remarks
Random-Aug (semantics-agnostic)	Large drop in GNN classification accuracy	Semantics destroyed
EPA-Aug (semantics-preserving)	GNN classification accuracy close to original	Semantics preserved
No explainer (full-graph random aug)	Poor performance	Lack of semantic protection
Varying label ratio (1–10%)	More labels → better explainer	Effective even at 1%

Key Findings

• On BA-2motifs, Random-Aug degrades GNN accuracy from near 100% to below 50%, while EPA-Aug maintains accuracy above 90%.
• t-SNE visualization confirms that EPA embedding distributions remain consistent with those of the original graphs, whereas Random-Aug causes significant distributional shift.
• EPA is plug-and-play and yields consistent improvements when combined with different contrastive learning frameworks (GraphCL/SimSiam).
• EPA remains effective even when as few as 1–5% of labeled graphs are used to train the explainer.

Highlights & Insights

• Bridging XAI and Data Augmentation: The connection is elegant — repurposing explainable AI techniques for data augmentation establishes a novel paradigm of "using explanations to protect semantics."
• Compelling Theoretical Guarantees: On the analytical model of BA-2motifs, the gap in classification error between semantics-preserving and semantics-agnostic encoders can be made arbitrarily large.
• Strong Practicality: EPA is plug-and-play and can be combined with any graph contrastive learning framework.

Limitations & Future Work

• Explainer quality depends on label quantity and GNN quality: If the pre-trained GNN is of poor quality, the explainer may identify incorrect semantic subgraphs.
• Additional pre-training cost: Training the GNN and explainer introduces an extra step compared to purely unsupervised methods.
• Assumption that semantics = subgraph: In some graphs, semantics may not reduce to simple subgraph structures (e.g., global attribute patterns).
• Evaluated only on graph-level tasks: Whether node-level contrastive learning can similarly benefit remains unexplored.

Related Work & Insights

• vs. GraphCL/JOAO: These methods rely on random augmentation without semantic preservation. EPA consistently outperforms them across all benchmarks.
• vs. AD-GCL: AD-GCL uses adversarial augmentation to learn what not to keep, whereas EPA uses an explainer to learn what to keep — the two approaches are complementary in direction.
• vs. ENGAGE: ENGAGE identifies important nodes in an unsupervised manner, but without label guidance it may capture structures that are not class-discriminative. EPA uses a small number of labels to ensure that the identified structures are genuinely semantically meaningful.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Bridging XAI and GRL is a fully novel direction, well-supported by both theory and experiments.
• Experimental Thoroughness: ⭐⭐⭐⭐ Six datasets, multiple frameworks, ablation studies, and visualizations.
• Writing Quality: ⭐⭐⭐⭐ Motivation is clear, the method is presented naturally, and the theoretical analysis adds value.
• Value: ⭐⭐⭐⭐ Provides a new perspective and tool for addressing "how to perform effective augmentation" in graph contrastive learning.

Related Papers

• [AAAI 2026] Let the Void Be Void: Robust Open-Set Semi-Supervised Learning via Selective Non-Alignment
• [AAAI 2026] HiLoMix: Robust High- and Low-Frequency Graph Learning Framework for Mixing Address Association
• [NeurIPS 2025] STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic Attention-Based Regional Masking
• [AAAI 2026] Self-Supervised Inductive Logic Programming
• [AAAI 2026] Improving Region Representation Learning from Urban Imagery with Noisy Long-Caption Supervision