Unsupervised Embedding Learning for Large-Scale Heterogeneous Networks Based on Metapath Graph Sampling.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Learning node embeddings in large-scale heterogeneous networks is a significant challenge in network representation.
• Existing methods often struggle with scalability and preserving semantic information in complex network structures.

Purpose of the Study:

• To propose an unsupervised embedding learning model, Large-scale Heterogeneous Graph Infomax (LHGI), for large-scale heterogeneous networks.
• To address the limitations of current methods in feature extraction and semantic information retention.

Main Methods:

• LHGI utilizes subgraph sampling guided by metapaths to compress the network while retaining semantic information.
• The model employs contrastive learning, maximizing mutual information between normal/negative node vectors and global graph vectors.
• This approach enables unsupervised training by optimizing the objective function without relying on labeled data.

Main Results:

• LHGI demonstrates superior feature extraction capabilities compared to baseline models on both medium-scale and large-scale unsupervised heterogeneous networks.
• The node embeddings generated by LHGI achieve enhanced performance in various downstream mining tasks.
• The subgraph sampling and contrastive learning strategies effectively capture network structure and semantics.

Conclusions:

• LHGI offers an effective unsupervised approach for learning node embeddings in large-scale heterogeneous networks.
• The model's ability to compress networks and retain semantic information leads to improved performance in downstream applications.
• LHGI advances the field of heterogeneous network embedding by providing a scalable and robust solution.