A Personalized Collaborative Filtering Recommendation System Based on Bi-Graph Embedding and Causal Reasoning.

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Existing recommendation systems struggle with feature bias and adapting to evolving user preferences.
• Graph embedding and collaborative filtering integration shows promise but faces limitations in personalization.

Purpose of the Study:

• To introduce RCKFM, a novel recommendation model designed to overcome feature bias and improve personalized recommendations.
• To enhance graph embedding technology and effectively model dynamic user interests over time.

Main Methods:

• Leveraged CoFM, TransR graph embedding, causal inference (backdoor tuning), KL divergence, and factorization machines.
• Employed TransR for diverse relationship types and causal inference to mitigate feature bias.
• Utilized KL divergence to predict and adapt to changes in user interests.

Main Results:

• RCKFM demonstrated superior performance on MovieLens-1M and Douban datasets.
• Achieved significant improvements (3.17%-6.81%) in precision, recall, NDCG, and hit rate for top-10 recommendations.
• Effectively addressed feature bias and captured evolving user preferences.

Conclusions:

• The proposed RCKFM model significantly enhances personalized recommendation accuracy.
• RCKFM offers a robust solution for feature bias and dynamic user interest modeling in recommendation systems.
• The findings highlight the potential impact of RCKFM in advancing the field of recommendation systems.