Large language model with retrieval semantics for cold-start recommendations

Area of Science:

• Recommender Systems
• Artificial Intelligence
• Machine Learning

Background:

• The item cold-start problem is a significant challenge in recommender systems.
• Existing methods suffer from semantic bias, semantic sparsity, and a behavioral gap due to lack of interactions.
• These limitations hinder the accurate recommendation of new items.

Purpose of the Study:

• To propose a novel method, Large Language Model with Retrieval Semantics (LaReS), to overcome the item cold-start problem.
• To reduce semantic bias by grounding cold items in reliable contexts using similar warm items.
• To enrich item semantics and bridge the interaction gap for improved cold-item recommendations.

Main Methods:

• Implemented self-retrieval contextualization to mitigate semantic bias by retrieving similar warm items.
• Utilized a large language model (LLM)-based semantic encoding strategy for dense, task-specific embeddings.
• Employed a joint preference learning framework with multi-objective optimization to bridge the interaction gap.

Main Results:

• LaReS demonstrates competitive performance in both cold-start and overall recommendation scenarios.
• The method significantly improves cold-start performance, with a 7.98% increase in P@5 on MovieLens.
• LaReS successfully grants cold items visibility comparable to established items.

Conclusions:

• LaReS effectively addresses the limitations of existing recommender systems in handling cold-start items.
• The proposed method enhances semantic understanding and bridges the interaction gap.
• LaReS offers a promising solution for improving the performance of recommender systems for new items.