Transfer learning from rating prediction to Top-k recommendation

Recommender system has made great strides in two major research fields, rating prediction and Top-k recommendation. In essence, rating prediction is a regression task, which aims to predict users scores on other items, while Top-k is a classification task selecting the items that users have the most potential to interact with. Both characterize users and items, but the optimization of parameters varies widely for their respective tasks. Inspired by the idea of transfer learning, we consider extracting the information learned from rating prediction models for serving for Top-k tasks. To this end, we propose a universal transfer model for recommender systems. The transfer model consists of two sub-components: quadruple-based Bayesian Converter (BC) and Prediction-based Multi-Layer Perceptron (PMLP). As the main part, BC is responsible for transforming the feature vectors extracted from the rating prediction model. Meanwhile, PMLP extracts the prediction ratings, constructs the prediction rating matrix, and uses multi-layer perceptron to enhance the final performance. On four benchmark datasets, we use the information extracted from the singular value decomposition plus plus (SVD++) model to demonstrate the effectiveness of BC-PMLP, comparing to classical and state-of-the-art baselines. We also conduct extra experiments to verify the utility of BC, and performance within different parameter values.