Improving Music Recommendation With Fine-Grained Content-Based Behavior Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Music streaming platforms generate extensive, often noisy, user listening histories.
• Generic sequential recommendation models face challenges in accurately modeling user interests due to this noisy data.
• Existing behavior retrieval techniques may not fully leverage the nuances of music content and temporal listening patterns.

Purpose of the Study:

• To introduce FactUBR, a fine-grained, content-based, and temporal user behavior retrieval framework for music recommendation.
• To address the limitations of generic sequential models in handling long and noisy listening histories.
• To enhance music recommendation by retrieving representative tracks that better reflect user preferences.

Main Methods:

• FactUBR integrates a reinforced content-based retrieval module (RCB) and a differentiable temporal-channel (DTC) retrieval module.
• RCB utilizes reinforcement learning to optimize retrieval decisions based on content differences, maximizing diversity and recommendation performance.
• DTC assesses fine-grained correlations between listening segments and candidate tracks, employing perturbed maximum techniques for optimization.

Main Results:

• FactUBR demonstrated significant improvements when enhancing various sequential recommendation models.
• The proposed framework outperformed existing state-of-the-art behavior retrieval techniques in experiments.
• Extensive evaluations were conducted on two public music recommendation benchmarks.

Conclusions:

• FactUBR offers a robust solution for improving music recommendations by effectively retrieving representative tracks.
• The framework's ability to exploit music content and temporal listening dynamics is crucial for enhanced performance.
• FactUBR represents a significant advancement in user behavior retrieval for sequential recommendation systems.