Learning Node Representations via Sketching the Generative Process With Events Benefits Link Prediction on...

Area of Science:

• Graph Representation Learning
• Network Science
• Machine Learning

Background:

• Heterogeneous Information Networks (HINs) are crucial for modeling complex real-world interactions.
• Representation learning on HINs generates compact embeddings for network analysis and machine learning.
• Existing methods often neglect the event-driven nature of HINs, limiting their ability to capture higher-order interactions and predict future links.

Purpose of the Study:

• To develop a novel method for representation learning on HINs that addresses the limitations of existing approaches.
• To enhance the capability of HIN embeddings to preserve higher-order interactions and predict potential links.
• To improve the performance of link prediction tasks in heterogeneous information networks.

Main Methods:

• Proposing Contrastive Learning via Events on Heterogeneous Information Networks (CLEH).
• Delineating the HIN generative process from local node structures to higher-order event structures.
• Designing an event-level contrastive learning procedure to capture higher-order node relations.
• Utilizing a normalizing flow model as an encoder to enhance embedding expressiveness.

Main Results:

• CLEH effectively captures higher-order relations among nodes by modeling the event generation process.
• The proposed method demonstrates significant superiority in link prediction tasks compared to existing baselines.
• Embeddings generated by CLEH are more expressive due to the use of a normalizing flow model.

Conclusions:

• CLEH offers a significant advancement in representation learning for Heterogeneous Information Networks.
• The event-centric approach effectively models higher-order interactions, leading to improved link prediction.
• CLEH provides a more comprehensive understanding of HINs by incorporating their generative dynamics.