Recommendation Methods With Deep Analysis On User Interests

In recent years,recommender systems have become one of the effective ways to solve the problem of information overload on user-generated contents.However,traditional recommendation methods are generally based on the collaborative filtering,most of which only consider the influence of user explicit interest on recommendation results.Deeply analyzing user behavior,we find that in a certain situation,the products and services recommended for users are not only affected by their own interest(explicit interest and implicit interest),but also influenced by the current situation.Therefore,we deeply mine and analyze the user behavior,user interest and the context where the user is in the online social network,and put forward four kinds of recommendation methods with deep analysis on user interest under the rich context-awareness.The main contributions of this thesis are as follows.1.A dual implicit mining-based latent friend recommendation is proposed.Deeply analyzing the influence of friend relationship on user activities,we find that the user implicit interest topic weight and implicit relationship between users in local topic clusters are very important to latent friend recommendation,based on which,a dual implicit mining-based latent friend recommendation model(DIM-LFR)is proposed.In the first implicit mining stage,the proposed ARAU-ATCI algorithm is used to identify the user interest topic,calculate the user interest topic weight and user implicit interest topic similarity.In the second implicit mining stage,the proposed WL-RWR algorithm is utilized to calculate the implied link similarity between users in local topic clusters.Then,the user implicit interest topic similarity and implicit link relationship similarity are combined together in a linear way,and top-N latent friends based on the dual implicit mining are recommended to the target user.Experimental results show that DIM-LFR outperforms the selected baselines.2.A deep membership and deep friendship awareness for social recommendation is proposed.As the influence of social relationships on recommended results is not carefully considered in existing social recommendations,first,the improved Jaccard similarity is used to calculate the deep membership similarity between users,and the two-hop restart random walk algorithm is used to compute the deep friendship similarity between users.Then,the deep similar membership and deep similar friendship are combined and merged into matrix factorization for social recommendation.Experimental results show that the proposed social recommendation is superior to the chosen baselines.3.A dual geo-social relationship and deep implicit topic mining for POI recommendation isproposed.Under online social networks,aiming at that most existing POI recommendations do not deeply and simultaneously consider user interest,social relationships based on the link relationship and common check-in behaviors,and deep implicit topics contained by POI reviews from users,a dual geo-social relationship and deep implicit interest topic similar relationship are used to model the user check-in behavior.SimRank similarity and Cosine similarity are adopted to extract the geo-social relationship between users,and the proposed RUA-TCP topic model is utilized to calculate the deep implicit interest similarity between users.Experimental results show that the proposed POI recommendation is better than the selected baselines.4.A POI true popularity and dual implicit trust mining for periodic POI recommendation is put forward.Under the location-based social networks,a periodic POI recommendation based on POI true popularity and dual implicit trust mining is proposed according to periodic changes of user interest.In each time slot,truly popular POIs and dual implicit trust mechanism(the implicit trust towards similar category experts and implicit trust to potential friends)are explored.First,the user check-in behaviors on POIs are divided according to the fixed time slot.Second,real popular POIs,similar category experts and the trusted friend set of the target user are mined in each time slot.Then,POI true popularity and dual implicit trust are utilized to constrain matrix factorization for periodic POI recommendation.Experimental results show that the proposed periodic POI recommendation is better than the chosen baselines.