Improved Diversity-Promoting Collaborative Metric Learning for Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Collaborative Metric Learning (CML) is a popular recommendation system (RS) technique.
• Existing CML methods use a unique user representation, which can cause bias with imbalanced item categories or diverse user interests.

Purpose of the Study:

• To propose Diversity-Promoting Collaborative Metric Learning (DPCML) to address preference bias in RS.
• To account for users' minority interests by introducing multiple user representations.

Main Methods:

• DPCML uses multiple representations per user, aggregating preferences via minimum distance across embeddings.
• Two assignment strategies and a Diversity Control Regularization Scheme (DCRS) are introduced.
• A novel sampling method is developed to address limitations of traditional negative sampling in CML.

Main Results:

• DPCML theoretically shows smaller generalization error compared to traditional CML.
• Experiments on benchmark datasets demonstrate the efficacy of DPCML.
• The proposed sampling method improves CML-based paradigms.

Conclusions:

• DPCML effectively mitigates preference bias in recommendation systems.
• The multi-vector representation and novel sampling enhance recommendation performance.
• DPCML offers a more robust approach for users with diverse interests.