Dual Consistency Constraint-Based Self-Supervised Representation Learning for Heterogeneous Graphs With Missing...

Area of Science:

• Graph representation learning
• Machine learning
• Data mining

Background:

• Heterogeneous graphs are complex data structures with diverse node and edge types.
• Missing attribute completion is crucial for leveraging heterogeneous graph data.
• Existing methods struggle with noisy attributes and data inconsistency across graph views.

Purpose of the Study:

• To develop a robust self-supervised representation learning method for heterogeneous graphs with missing attributes.
• To address noise propagation and data inconsistency issues in current approaches.
• To enhance the accuracy and effectiveness of attribute completion in heterogeneous graphs.

Main Methods:

• Proposes a dual consistency constraint-based self-supervised learning framework.
• Incorporates representation completion and within-view consistency loss to handle missing attributes and noise.
• Utilizes cross-view consistency loss to ensure data consistency across augmented graph views.
• Reconstructs masked data to mitigate information loss during the learning process.

Main Results:

• The proposed method effectively filters out noise and inaccurate information.
• Achieves discriminative representation learning for heterogeneous graphs with missing attributes.
• Demonstrates superior performance on various downstream tasks compared to existing methods.

Conclusions:

• The dual consistency constraint method offers a significant advancement in heterogeneous graph representation learning.
• This approach provides a robust solution for handling missing attributes and noisy data.
• The findings have implications for various applications utilizing heterogeneous graph data.