HeteEdgeWalk: A Heterogeneous Edge Memory Random Walk for Heterogeneous Information Network Embedding.

Area of Science:

• Graph Representation Learning
• Network Science
• Data Mining

Background:

• Traditional Heterogeneous Information Network (HIN) embedding methods rely on meta-paths to guide random walks, aiming to mitigate biases towards high-order nodes.
• The performance of meta-path-based methods is heavily contingent on the suitability of predefined meta-paths, which often necessitates domain expertise and struggles to capture complex HIN structures.
• Existing methods can ignore valuable heterogeneous information not specified within the chosen meta-paths.

Purpose of the Study:

• To propose HeteEdgeWalk, a novel meta-path-free approach for HIN embedding.
• To develop a dynamically adjusted bidirectional edge-sampling walk strategy for more balanced and comprehensive network structure sampling.
• To evaluate the effectiveness of HeteEdgeWalk in capturing semantic information from HINs for downstream tasks.

Main Methods:

• Introduced HeteEdgeWalk, a method that avoids the use of meta-paths for HIN embedding.
• Designed a dynamically adjusted bidirectional edge-sampling walk strategy.
• Incorporated edge sampling and storage of recently selected edge types to improve network sampling.

Main Results:

• Node classification experiments showed a maximum performance improvement of 2% compared to baseline methods.
• Clustering experiments achieved at least a 0.6% performance improvement over baselines.
• Demonstrated the method's superiority in effectively capturing semantic information from complex HINs.

Conclusions:

• HeteEdgeWalk offers a superior alternative to meta-path-guided methods for HIN embedding.
• The proposed edge-sampling strategy effectively addresses limitations of traditional random walks in HINs.
• The meta-path-free approach simplifies the process and enhances the representation learning of HINs.