FTF-ER: Feature-Topology Fusion-Based Experience Replay Method for Continual Graph Learning (2407.19429v3)

Abstract: Continual graph learning (CGL) is an important and challenging task that aims to extend static GNNs to dynamic task flow scenarios. As one of the mainstream CGL methods, the experience replay (ER) method receives widespread attention due to its superior performance. However, existing ER methods focus on identifying samples by feature significance or topological relevance, which limits their utilization of comprehensive graph data. In addition, the topology-based ER methods only consider local topological information and add neighboring nodes to the buffer, which ignores the global topological information and increases memory overhead. To bridge these gaps, we propose a novel method called Feature-Topology Fusion-based Experience Replay (FTF-ER) to effectively mitigate the catastrophic forgetting issue with enhanced efficiency. Specifically, from an overall perspective to maximize the utilization of the entire graph data, we propose a highly complementary approach including both feature and global topological information, which can significantly improve the effectiveness of the sampled nodes. Moreover, to further utilize global topological information, we propose Hodge Potential Score (HPS) as a novel module to calculate the topological importance of nodes. HPS derives a global node ranking via Hodge decomposition on graphs, providing more accurate global topological information compared to neighbor sampling. By excluding neighbor sampling, HPS significantly reduces buffer storage costs for acquiring topological information and simultaneously decreases training time. Compared with state-of-the-art methods, FTF-ER achieves a significant improvement of 3.6% in AA and 7.1% in AF on the OGB-Arxiv dataset, demonstrating its superior performance in the class-incremental learning setting.

• The paper introduces FTF-ER, which fuses feature and topology data in an innovative experience replay mechanism to address catastrophic forgetting in graph neural networks.
• It utilizes a dynamic buffer and a mixing strategy, integrating established GNN architectures and the Hodge decomposition theorem to select key nodes and subgraphs.
• Empirical results on multiple datasets demonstrate that FTF-ER significantly outperforms existing methods like GEM and ER-GNN, highlighting its robustness in sequential learning.

Overview of FTF-ER: Feature-Topology Fusion-Based Experience Replay Method for Continual Graph Learning

This paper presents a novel approach to continual graph learning through a methodology termed Feature-Topology Fusion-Based Experience Replay (FTF-ER). The proposed framework addresses challenges in Graph Neural Networks (GNNs), especially when applied to continual learning tasks. The paper aims to enhance model performance across sequential tasks by developing a strategy that integrates both feature and topological information effectively.

Core Contributions and Methodology

The FTF-ER framework is designed to deal with a sequence of tasks denoted by $$\mathcal{T}=\{\mathcal{T}_1, \mathcal{T}_2, \ldots, \mathcal{T}_K\}$$. The key innovation lies in the use of an experience buffer $\mathcal{B}$ that dynamically captures essential information from previously encountered tasks. This buffer is used to replay experiences that help in retaining crucial information while learning new tasks, addressing the common challenge of catastrophic forgetting in neural networks.

1. Experience Replay Mechanism:
   • The method maintains an experience buffer, $\mathcal{B}$, which is continuously updated with representative nodes and subgraphs from each task. This selection is informed by a fusion of topological and feature relevance.
   • A mixing strategy, $\mathbf{S}^{mix}$, guides the selection of nodes from the training dataset to populate the buffer, thereby ensuring diversity and representativeness in retained examples.
2. Graph Neural Networks as Framework Backbones:
   • The paper utilizes well-established GNN architectures including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Isomorphism Networks (GIN) as backbones. These networks are leveraged to transform and aggregate neighborhood information at each learning layer.
3. Application of the Hodge Decomposition Theorem:
   • A mathematical foundation underpinning the approach is the application of the Hodge Decomposition on graphs, adapted from its classical form on Riemannian manifolds. This theorem allows node importance to be quantified based on their topological positions in the graph, contributing to the feature-topology fusion strategy.

Numerical Results and Significance

The paper reports substantial empirical results across various datasets, including Amazon Computers, Corafull, OGB-Arxiv, and Reddit. The proposed FTF-ER method consistently outperforms existing continual learning strategies such as GEM, ER-GNN, and SSM. Particularly notable is its robustness across different datasets, with the performance evaluated through accuracy metrics that highlight its adaptability to diverse tasks.

An extensive sensitivity analysis on the hyper-parameter $\beta$ further illustrates the versatility and reliability of FTF-ER. The analysis reveals that while the method is sensitive to $\beta$, optimal configurations consistently reside around mid-range values, thus indicating a stable and predictable tuning behavior.

Implications and Future Directions

The implications of this research are both practical and theoretical. Practically, FTF-ER provides a clear route to efficiently manage sequential task learning in GNNs, significantly mitigating forgetting effects while maximizing task retention. The fusion strategy innovatively combines feature discipline with topological insight, which can be tailored for broader applications, particularly in domains where graph topology plays a critical role, such as social network analysis and pharmacogenomics.

Theoretically, the application of Hodge decomposition onto graph structures offers a new analytical lens for assessing node significance, suggesting avenues for further exploration in topologically-informed machine learning strategies.

Conclusion

The introduction of FTF-ER constitutes an impactful increment in the field of continual graph learning. By seamlessly integrating feature and topological insights via an experience replay paradigm, the authors provide a comprehensive strategy that addresses continual learning's most pressing issues. Future research may focus on extending this method to accommodate larger and more complex graph datasets in real-world scenarios, exploring potential intersections with reinforcement learning, and further refining the theoretical underpinnings of graph learning architectures.