HetDualCL: Dual-encoder contrastive learning for heterogeneous graphs

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Heterogeneous graphs (HGs) model complex systems with diverse node and relation types.
• Current representation learning methods for HGs often require extensive labeled data and struggle with integrating information from multiple sources.
• Existing self-supervised techniques are frequently constrained by single-perspective information processing and limited encoder architectures, impeding the comprehensive integration of multi-granularity semantic information.

Purpose of the Study:

• To address the limitations of current methods in heterogeneous graph representation learning.
• To propose a novel framework that systematically integrates local topology and long-range semantics.
• To improve the discriminative power of node representations in heterogeneous graphs.

Main Methods:

• Introduced HetDualCL, a dual-encoder contrastive learning framework.
• Developed an enhanced Graph Neural Network (GNN) encoder for robust local topological modeling.
• Designed a Gated Causal Convolutional (GCC) encoder to capture multi-hop semantic dependencies.
• Employed a cross-view contrastive loss to collaboratively align and optimize local and semantic views.

Main Results:

• HetDualCL effectively integrates local topology and multi-hop semantics.
• The framework learns highly discriminative node representations.
• Achieved superior performance in node classification and clustering tasks across four benchmark datasets.

Conclusions:

• HetDualCL offers a powerful approach for heterogeneous graph representation learning.
• The dual-encoder strategy successfully bridges the gap in integrating multi-granularity semantic information.
• The proposed method demonstrates significant improvements over existing techniques for tasks involving complex graph structures.