Personalized Latent Structure Learning for Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• User behavior in recommender systems is driven by latent user-item factors.
• Current methods focus on latent factor disentanglement but neglect their dependencies.
• Uncovering latent factor interactions is crucial for robust recommendations.

Purpose of the Study:

• To investigate joint disentanglement of user-item latent factors and their dependencies (latent structure learning).
• To address challenges in recommendation-specific latent structure learning, including user subjectivity and data inaccessibility.
• To propose a personalized latent structure learning framework for recommendation.

Main Methods:

• Analyzing latent structure learning from a causal perspective, ensuring structure acyclicity and dependency constraints.
• Developing PlanRec, a framework incorporating differentiable Reconstruction, Dependency, and Acyclicity regularizations.
• Implementing Personalized Structure Learning (PSL) and uncertainty estimation for adaptive personalization.

Main Results:

• PlanRec effectively discovers shared and personalized latent structures.
• The framework successfully balances shared knowledge and personalization using uncertainty estimation.
• Experiments on MovieLens, Amazon, and Alipay datasets validate PlanRec's effectiveness.

Conclusions:

• Latent structure learning is essential for advancing recommender systems.
• PlanRec offers a robust solution for personalized latent structure learning.
• The proposed method enhances recommendation effectiveness by adapting to individual user needs.