Personalized Recommendation That Integrated Social Information

With the rapid growth and wide application of social networks,while enjoying the new life brought by science and technology,people are facing the dilemma of "information overload".Recommendation system is considered to be one of the most effective ways to solve information overload,and has attracted wide attention from industry and academia.As a socially bonded group,one of human beings' basic attributes is Social.Social is also an essential element for the development of modern Internet applications.Thanks to the continuous improvement of online social functions and the large supply of social information,the integration of social information to enhance the effectiveness of personalized recommendation has become a hot research topic in the field of recommendation in recent years.Although researchers have achieved lots of research results on related area,there are still some deficiencies in the accuracy and diversity of the existing research results.Therefore,aiming at the characteristics of specific networks and the shortcomings of existing research work,this paper deeply studies personalized recommendation that integrates social information from three aspects,and achieves the following results:First,aiming at the accuracy of user recommendation in Uni-directional social networks,this paper studies personalized recommendation based on fusion of following information,and proposes a new user recommendation framework called UIS-MF.Filtering the targeted users is a demanding and key task on Uni-directional social networks.Related works prefer to utilize follower-followee relations for recommendation.However,a ma-jor problem of these methods is that they assume the motives of every follower-followee user pairs are the same,does not able to separate the interest motivation and social motivation,and this leads to the coarse user following preferences inferring and unsatisfied recommending ac-curacy.In this regard,this paper propose a two stage recommendation framework UIS-MF.Firstly,a unified probabilistic topic model UIS-LDA is proposed on follower-followee relations for discovering interest topics and social topics of users.The innovative point of this model is to introduce the Generalized Polya Urn model into sampling process to effectively guide the generation of the two types of topics.It is also the first topic model to jointly model users' in-terest preferences and social preferences.Secondly,a community-based method is proposed for user recommendation,it organizes social communities and interest communities based on the estimation of topics obtained from UIS-LDA,and then performs Matrix Factorization method on each community to generate N most likely followees who have similar interest and close social connection relevant to a target user.Systematic experiments on Twitter,Sina Weibo and Epinions datasets have not only revealed the significant effect of UIS-LDA model for the ex-traction of interest and social topics of users in improving recommending accuracy,but also demonstrated the advantage of our UIS-MF over competitive baselines on all by large margins.Also,the community-based method alleviates the problem of data sparsity to a certain extent.Second,regarding to the accuracy of recommendation in online trust network,this pa-per studies the personalized recommendation by combining trust information and proposes two better recommendation methods incorporating dual roles influence called BPRDR and FSDR.Beside,a recommendation method based on trust and distrust relationship is proposed based on BPRDR framework called BPRTaD.This paper conducts an in-depth analysis on Epinions,Ciao,and FilmTrust datasets aiming at offering fundamental support to the trust-based research for item recommendation.We find that a user's selection of an item is influenced not only by her trustees but also by her trusters.In this regard,this paper leverages this "dual roles influence" to derive two more accurate Matrix Factorization based ranking models BPRDR and FSDR,respectively.In more detail,the first BPRDR model performs three pairwise preferences comparisons under the Bayesian Personal Ranking framework,considering the dual roles influence in its ranking assumptions.The sec-ond FSDR adopts the factored similarity of the implicit vector of the feature matrix to measure the " dual roles influence",and combines the user-user similarity and the item-item similarity to calculate the ranking preference score for user.Finally,both methods generate recommendation lists containing N items for target users using ranking preference scores.Extensive experiments on three datasets show that it is essential to consider the dual roles influence when generating Top-N item recommendation.In addition,aiming at verifying the influence of distrust relations on item ranking,this paper develops a novel model BPRTaD based on BPRDR.Unlike BPRDR,BPRTaD considers the "dual roles influence" coming from users' trustees and distrustees.Nev-ertheless,BPRTaD and BPRDR have similar pairwise hypothetical structures,optimization rules and recommendation methods.Experiments show that BPRTaD is superior to other compara-tive methods.It shows that the incorporation of distrust relationship can effectively improve the accuracy of recommendation,and this also brings beneficial inspiration to the new research of personalized recommendation in trust networksThird,aiming at the dilemma of accuracy and diversity of online recommendation,this pa-per studies the integration of friend information and curiosity for personalized recommendation,and proposes a curiosity-driven recommendation framework namely CdRFOveremphasizing accuracy in recommendation systems can make users feel bored,while overemphasizing diversity can make users feel uncomfortable.Aiming at the dilemma,a highly innovative Stimulus-evoked Curiosity Model SeCM is presented in this paper,which builds on top of the three well-known theories in psychology:Curiosity-drive,Social Conflict,and Inter-mediate Arousal Potential.SeCM first estimates the stimulus intensity appearing on each item for each user.Specifically,a combination of novel and conflict stimulus is applied in this work Among them,friendship is used to calculate the intensity of conflict stimulus for the first time SeCM then models personalized curiosity among the calculated stimulus intensities by using the Wundt curve.SeCM's innovation is that it is the first time to quantify the intensity of con-flict stimulus from the perspective of friend-item feedback.It is also the first time to synthesize the curiosity mechanism induced by two stimuli,and construct the Wundt curve with learning ability for the first time.Finally,this work proposes a highly innovative Curiosity-drive Recom-mendation Framework CdRF which incorporates SeCM together with a basic accuracy-oriented Matrix Factorization algorithm via weighted Borda count to balance the accuracy and diversity for Top-N recommendation.For the target user,the output of CdRF is a ranked list of N items which are both relevant and highly curiousness.Experiments are conducted against 14 com-petitors(specifications)using movie datasets from MovieLens-1M and-100K to evaluate the performance of each component of SeCM as well as the whole framework CdRF.The results reveal that SeCM can significantly enhance diversity,and CdRF can flexibly generate diversi-fied recommendation while keeping the recommendation accuracy relatively high,which better alleviates the dilemma of accuracy and diversity.