Continual Augmentation for Graph Contrastive Learning
- 주제(키워드) 그래프 대조학습 , 그래프 증강
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 신현정
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035703
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Graph Contrastive Learning (GCL) has emerged as a powerful paradigm for self- supervised representation learning. However, the quality of learned representations depends heavily on the design of graph augmentations. Existing augmentation strategies rely on random or topology-based bulk perturbations. Such approaches fail to capture the dependencies between edges and disregard node-feature semantics, potentially destroying core structural patterns and causing semantic drift. To address these limitations, we propose Continual Augmentation for Graph Contrastive Learning (CA-GCL), a novel framework that reformulates graph augmentation as a sequential decision-making problem. We design a learnable agent that modifies the graph structure step by step and optimize agent via reinforcement learning. To guide the agent, we introduce a semantics-aware reward derived from the Graph Diffused Feature Distance (GDFD). Our theoretical analysis shows that GDFD decomposes into complementary topological and node-feature terms, providing an explicit criterion for preserving both semantics during augmentation. Empirical results on graph classification show that CA-GCL outperforms baselines by generating contrastive views that are semantically consistent.
more목차
1. Introduction 1
2. Preliminaries 5
2.1 Notations 6
2.2 Graph Neural Networks 6
2.3 Graph Contrastive Learning 7
2.4 Reinforcement Learning 8
2.5 Graph Spectrum 9
3. Related Works 11
3.1 Classical Graph Contrastive Learning 12
3.2 Graph Contrastive Learning with Learnable Augmentation 12
4. Proposed Method 14
4.1 Continual Augmentation Scheme 16
4.2 Parameterized Policy Network Design 17
4.2.1 State Encoding and Edge Representation 18
4.2.2 Edge dropping probability 18
4.3 Reward Design 19
4.3.1 Graph Diffusion Wasserstein Distance 20
4.3.2 Graph Diffused Feature Distance 20
4.3.3 Theoretical Analysis of Semantics Preservation 21
4.3.4 Step-wise Reward Formulation & Optimization 26
4.4 Extensibility to Other Augmentation Strategies 26
5. Experiments 29
5.1 Qualitative Analysis 30
5.1.1 Synthetic Datasets 30
5.1.2 Experimental Setup (Augmentation) 32
5.1.3 Case Study on Synthetic Graphs – Visualization 32
5.2 Quantitative Analysis 35
5.2.1 Benchmark Datasets – Graph Classification 35
5.2.2 Experimental Setting 36
5.2.3 Graph Classification Results 37
5.3 Ablation Analysis 40
5.4 Additional Analysis 41
6. Conclusion 43
References 45
