Class-aware graph Siamese representation learning

Chengcheng Xu

doi:10.1016/j.neucom.2024.129209

2025-03-01

PRE

DOI：10.1016/j.neucom.2024.129209

原文链接

原文求助

摘要

En 中文

Currently, two issues exist in the field of graph Siamese representation learning. First, the strategies for positive sample selection often impose strict constraints on the candidate set, which may result in critical class information being overlooked. Second, traditional data augmentation strategies often alter the topological structure of graphs, which may impact the representations of nodes. Therefore, we propose a class-aware graph Siamese representation learning method named CGSRL to address these issues. Inspired by the Jaccard coefficient, we design a node structural similarity calculation method that considers neighborhood information and average degree. In the selection of positive samples, we comprehensively considered both feature similarity and structural similarity, allowing for the selection of positive samples belonging to the same class over a wider range. Additionally, we craft a feature augmentation method that preserves the raw class information. Specifically, CGSRL enforces a discriminator to extract high-level semantics representing class information from the raw features through carefully designed consistency constraints. In turn, a generator synthesizes augmented features with the same class information as the raw features based on these high-level semantics. To validate the superiority of our proposed method, we conducted experiments in node classification and node clustering on five real-world graph datasets. The results consistently demonstrate that our proposed CGSRL method outperforms state-of-the-art baselines.

Keyword:

Graph Siamese representation learning

Positive sample selection

Graph data augmentation

Class consistency