Transfer learning for collaborative recommendation with biased and unbiased data

Abstract

In a recommender system, a user's interaction is often biased by the items' displaying positions and popularity, as well as the user's self-selection. Most existing recommendation models are built using such a biased user-system interaction data alone. In this paper, we introduce an additional specially collected unbiased data, and then have a new problem called collaborative recommendation with biased and unbiased data.We first formalize the studied problem and list three challenges, including the bias challenge, the heterogeneity challenge and the unbalance challenge. Then we propose a novel transfer learning-based AI solution, i.e., transfer via joint reconstruction (TJR), to achieve knowledge transfer and sharing between the biased data and unbiased data. Specifically, in our TJR, we use two different models to extract the users' preferences and bias information, and then refine the prediction via the latent features containing the bias information in order to obtain a more accurate and unbiased recommendation. We further integrate the two data by reconstructing their interaction in a joint learning manner. Moreover, in order to better address the unbalance challenge, we introduce a bias regularization term and integrate bidirectional knowledge distillation. Finally, we adopt four representative methods, i.e., variational autoencoders, matrix factorization, neural collaborative filtering and graph convolution network, as the backbone models of our TJR and conduct extensive empirical studies on three public datasets, showcasing the effectiveness of our transfer learning solution over some very competitive baselines.