CoarSAS2hvec: Heterogeneous Information Network Embedding with Balanced Network Sampling.

Area of Science:

• Data Science
• Network Analysis
• Machine Learning

Background:

• Heterogeneous Information Network (HIN) embedding is crucial for analyzing complex relationships.
• Existing random-walk methods suffer from imbalanced sampling due to hub nodes.
• This imbalance limits the effectiveness of HIN embedding techniques.

Purpose of the Study:

• To propose a novel HIN embedding method, CoarSAS2hvec, that overcomes sampling limitations.
• To enhance the quality of node representations in HINs.
• To improve performance in downstream tasks like node classification and community detection.

Main Methods:

• Utilized self-avoiding short sequence sampling with HIN coarsening (CoarSAS) for richer context collection.
• Employed an optimized loss function to refine HIN structure embedding.
• Evaluated CoarSAS2hvec against nine other methods on four real-world datasets.

Main Results:

• CoarSAS2hvec demonstrated superior performance in node classification and community detection.
• Information-theoretic analysis confirmed CoarSAS captures richer network information than other methods.
• The proposed method significantly improves embedding quality even with traditional loss functions.

Conclusions:

• CoarSAS2hvec effectively addresses the imbalanced sampling problem in random-walk-based HIN embedding.
• The CoarSAS procedure enhances information capture, leading to better embedding performance.
• This research offers a new perspective for improving HIN analysis.