Multi-Granularity Graph Contrastive Learning Framework via Granular-Ball on Heterogeneous Graphs

Area of Science:

• Graph representation learning
• Self-supervised learning
• Machine learning

Background:

• Graph Contrastive Learning (GCL) is a popular self-supervised learning (SSL) technique.
• Mainstream GCL methods often overlook graph topology and multi-granularity characteristics, potentially damaging structural integrity through single augmentation schemes.
• Heterogeneous graphs present unique challenges where interconnected nodes may share semantic information despite differing classes.

Purpose of the Study:

• To propose a novel multi-granularity graph contrastive learning framework (GBGCL) specifically designed for heterogeneous graphs.
• To address the limitations of existing GCL methods in handling graph topology and multi-granularity.
• To enhance node representation learning by integrating feature and semantic information.

Main Methods:

• Developed an adaptive granular-ball augmentation strategy to identify multi-granularity homogeneous regions within heterogeneous graphs.
• Defined positive pairs as nodes within the same granular-ball and negative pairs as nodes from different granular-balls.
• Integrated original node features with semantic information of similar nodes in the feature space for joint optimization.

Main Results:

• The proposed GBGCL framework demonstrates unique advantages in learning node representations on heterogeneous graphs.
• The adaptive granular-ball augmentation effectively captures multi-granularity characteristics.
• Joint optimization of feature and semantic information leads to improved node representations.

Conclusions:

• GBGCL offers a powerful approach for self-supervised learning on heterogeneous graphs.
• The framework successfully leverages multi-granularity information and semantic similarities.
• This method shows significant potential for advancing graph representation learning in complex network structures.