Similarity-based Regularized Latent Feature Model for Link Prediction in Bipartite Networks

Area of Science:

• Data Mining
• Network Analysis
• Machine Learning

Background:

• Complex real-world systems are often modeled as bipartite networks.
• Link prediction in these networks is crucial for recommendation systems and understanding network evolution.
• Existing methods may not fully capture local network characteristics.

Purpose of the Study:

• To propose a novel framework for link prediction in bipartite networks.
• To enhance the accuracy and stability of link prediction by integrating structural and feature-based approaches.
• To address limitations in current methods by considering local network geometry.

Main Methods:

• Developed Similarity Regularized Nonnegative Matrix Factorization (SRNMF).
• Incorporated similarity-based structure and latent feature models.
• Utilized an iterative gradient descent scheme to optimize the objective function.
• Constructed a similarity matrix to encode network geometrical information.

Main Results:

• SRNMF demonstrated competitive and stable performance across various real-world bipartite networks.
• The proposed framework outperformed existing state-of-the-art link prediction methods.
• Explicitly considering local characteristics and geometrical information improved prediction accuracy.

Conclusions:

• SRNMF offers a robust and effective approach for link prediction in bipartite networks.
• The method provides significant improvements for applications like recommendation systems.
• This framework advances the field of data mining and complex network analysis.