Self-Supervised Heterogeneous Graph Neural Network with Multi-scale Meta-Path Contrastive Learning

Overview

Heterogeneous graph neural networks (HGNNs) present significant challenges in simultaneously capturing local neighborhood structure and global relational patterns. This work introduces HMMC, a self-supervised framework that leverages multi-scale meta-path embeddings and contrastive learning to improve heterogeneous graph representation learning. The framework addresses fundamental limitations in existing methods where meta-path selection involves trade-offs between representation completeness and noise introduction.

Methods and approach

HMMC integrates two primary technical components. The multi-scale meta-path embedding mechanism operates concurrently across different meta-path lengths to capture structural information at varying granularities, mitigating issues where excessively short paths provide insufficient context and excessively long paths introduce noise. The framework employs a cross-view self-supervised contrastive learning approach that optimizes representations across multiple perspectives of the heterogeneous graph structure. A novel star-shaped contrastive loss function is introduced to address the degradation caused by noisy negative samples in conventional contrastive learning paradigms. This loss exploits center-neighborhood structural dependencies as supervisory signals, maintaining representation consistency for positive pairs while preventing over-smoothing effects that arise from embedding convergence among similar nodes.

Key Findings

Experimental validation across multiple public heterogeneous graph datasets demonstrates quantitative improvements over established baseline methods, with performance gains ranging from 0.5% to 4.1% across different benchmarks. The results indicate enhanced representation power, robustness, and generalization capability relative to state-of-the-art approaches. The multi-scale meta-path mechanism successfully resolves the inherent tension between local structural granularity and global semantic consistency in heterogeneous graph representation learning.

Implications

The proposed methodology advances heterogeneous graph representation learning by providing a unified framework that simultaneously addresses multiple technical challenges in the domain. The star-shaped contrastive loss function presents a generalizable approach to mitigating negative sample noise in self-supervised settings beyond heterogeneous graphs. The demonstrated performance improvements suggest the framework has substantial applicability to downstream tasks requiring robust heterogeneous graph representations.