A Study Of Collaborative Filtering Recommendation Algorithms Based On Multi-source Information

The rapid development of web technology brings new problems and challenges to the recommender systems.On the one hand,the traditional collaborative filtering approaches have been difficult to meet user's personalized recommendation needs.On the other hand,the massive data brought by the web technology provides more useful information for the recommendation algorithms,how to extract features from these information and alleviate the three classic problems of sparsity,cold start and temporal dynamics in recommender systems,and effectively improve the recommendation quality are the hot topics in current recommendation algorithm researches.This paper does three aspects of in-depth research on the recommendation problem based on multi-source data:(1)Aiming at the three problems of quick changes of user interest,the lack of effective auxiliary information and repeated purchase in the context of physical store,a set of user long-term and short-term interests fused recommendation algorithm is proposed.The fusion of graph computing and matrix decomposition,bayesian personalized recommendation and tensor decomposition solves the rating prediction,item ranking and new time window based prediction problem respectively.User indoor trajectory information is exploited and mined to extract user behavior feature as the auxiliary information,and heuristic rules are designed to solve the problem of repeated purchase.(2)Aiming at the lack of effective multi-source information fusion mechanism in existing researches,an improved factorization machine model is proposed.User's social relation and review information are extracted and fused by expanding the input vector domains,solving the problem of fusion machanisum for these two important types of auxiliary information.The objective function is improved by adding social regularization term and vector internal domain regularization term for the consideration of actual meaning and the structure of input vector,which alleviates the problem of over-fitting.Furthermore,a deep feature fusion model is proposed and deep learning methods are used to reconstruct the input vector and fuse features for learning to perform rating prediction.This solves the problem that deep learning methods deal with categorical multi-domain discrete data.(3)Aiming at the problem of insufficient expression ability by using item similarity aggregation model,a more interpretive and expressive hybrid factorized user and item similarity aggregation method is proposed.Aiming at the personalized recommendation problem dominated by user short-term preference in e-commerce scenarios,a global scoring function with time-decay factor is proposed to model multi-type implicit feedback data,and a constrained sampling method is proposed to process negative feedback in implicit feedback.Aiming at the problem of collaborative modeling between deterministic feedback and non-deterministic feedback in implicit feedback information,it is proposed that the deterministic implicit feedback and non-deterministic implicit feedback information be aggregated and calculated by using two coefficient matrices.The proposed method solves the problem of cooperative modeling using two types of of implicit feedback information,and regularization term is added to the objective function,which makes the model have better expressive ability and practical significance.Extensive experiments on real datasets show that the proposed methods have a higher recommendation quality than the classical and state-of-the-art approaches.