A Bayesian matrix factorization model for dynamic user embedding in recommender system

1 Introduction The main idea of recommender system is how to learn accurate users'embeddings from behavior data[1].Each dimension of users'embeddings can reflect the interests of users in different potential aspects.In real-world scenarios,users'interests are drifting over time,which brings a chal-lenge to learn accurate dynamic users'embeddings.Recently,various time-aware recommendation methods have been proposed to tackle this problem by modeling the evolution process of users'interests[2-4].However,they usually assume that users'embeddings drift with the same range on all dimensions.In practice,users'embeddings should change diversely on different dimensions over time.Specifically,for the rapidly changing interests of the users,the corresponding dimensions should change significantly.On the contrary,the dimensions representing stable interests may change slightly.