Knowledge Transfer via Pre-training for Recommendation: A Review and Prospect

Area of Science:

• Artificial Intelligence
• Machine Learning
• Information Retrieval

Background:

• Recommender systems frequently encounter data sparsity, particularly the cold-start problem, hindering accurate user recommendations.
• Pre-trained models offer a promising solution by leveraging knowledge transfer across different domains and tasks.

Purpose of the Study:

• To survey the application of pre-trained models in recommender systems.
• To experimentally validate the benefits of pre-training for mitigating data sparsity.
• To identify future research avenues in this domain.

Main Methods:

• Comprehensive literature review of pre-trained models applied to recommender systems.
• Empirical evaluation demonstrating the performance improvements and data sparsity alleviation.
• Analysis of current trends and challenges to propose future research directions.

Main Results:

• Pre-trained models significantly enhance recommender system performance by overcoming data sparsity.
• Knowledge transfer capabilities of pre-trained models are crucial for improving recommendation accuracy.
• Experimental results confirm the practical benefits of integrating pre-training into recommender systems.

Conclusions:

• Pre-trained models are a powerful tool for addressing data sparsity in recommender systems.
• Further research into novel pre-training strategies and architectures holds significant potential.
• The availability of source code aims to foster reproducible research and accelerate advancements.