Abstract: In order to further improve the performance of tag-aware personalized recommendation with implicit feedbacks,aiming at the problems of the redundancy,ambiguity of tagging information and the sparsity and imbalance of implicit feedbacks,this paper proposed a personalized recommendation method based on fine-grained preference assumption and augmented weighted matrix factorization.First,one kind of candidate items that the target user may prefer are mined by leveraging its neighbor user,which are preferred by neighbor users which have not been selected by the target user.Thus,a type of fine-grained preference relationship among three kinds of items for target users is obtained,i.e.,observed item>candidate item>other unobserved data.This kind of operation can help to alleviate the sparsity and imbalance problem.Then,the deep learning method is used to extract the in-depth semantic features from tag space.In this way,representations of users’ profiles become more abstract and advanced,and user neighbors are obtained based on the in-depth semantic features.Afterwards,a revised weighted matrix factorization model is formulated based on the fine-grained preference relationship for personalized recommendation.And a fast eALS algorithm is used for model optimization in terms of low time complexity.Experiments on real-world datasets show that the proposed method outperforms competing methods on several evaluation metrics,including Pre@5,NDCG@5,MRR.The three indicators are respectively increased by 9%,8%,and 9%,which indicates the effectiveness of the proposed methods.

Key words: Deep learning, Fine-grained preference, Implicit feedback, Tag-aware recommendation, Weighted matrix factorization