Transfer learning from rating prediction to Top-k recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommender systems excel in rating prediction (regression) and Top-k recommendation (classification).
• Parameter optimization differs significantly between these two tasks.
• Transfer learning offers a potential method to bridge this gap by reusing learned information.

Purpose of the Study:

• To propose a universal transfer model for recommender systems.
• To extract and utilize information from rating prediction models for Top-k recommendation tasks.
• To enhance the performance of Top-k recommendation by transferring learned features.

Main Methods:

• Developed a universal transfer model comprising a Bayesian Converter (BC) and a Prediction-based Multi-Layer Perceptron (PMLP).
• BC transforms feature vectors from rating prediction models.
• PMLP extracts prediction ratings, constructs a rating matrix, and applies multi-layer perceptron for performance enhancement.

Main Results:

• Demonstrated the effectiveness of the BC-PMLP model on four benchmark datasets.
• Utilized information extracted from the Singular Value Decomposition plus plus (SVD++) model.
• Outperformed classical and state-of-the-art baseline methods in Top-k recommendation tasks.

Conclusions:

• The proposed BC-PMLP model successfully transfers knowledge from rating prediction to Top-k recommendation.
• The model shows significant improvements in recommendation performance.
• Further experiments confirmed the utility of BC and the impact of parameter variations.