Personalized Recommendation for Cold-Start Users via Cluster-Level Latent Feature Model

Abstract

Making accurate and personalized recommendations for the cold-start users is a significant yet challenging problem in the cross-domain recommendation systems. In collaborative filtering methods, personalised recommendations are concluded from user history data. Therefore, a large volume is necessary for an accurate recommendation system. In this paper, we extend Wang et al.'s work on latent feature mapping model in single auxiliary domain into two auxiliary domains for the cold-start users who are in the target domain. First of all, in order to better describe users in the sparse domains, we make full use of K-means algorithm to extract user's feature between two auxiliary domains and handle the rating matrix in the target domain with matrix factorization by combining user similarities. Next, to perform knowledge transfer across different domains, we employ a neighborhood based multi-layer perceptron (MLP) approach to learn feature mapping function of users. Finally, the preference of the cold-start user in the target domain could be predicted based on the mapping function and his/her latent factors in the auxiliary domain. In addition, experimental results on user data obtained from Amazon platform reveal the edge of our new model against other relevant methods.