Collaborative Filtering Algorithm Based On User's Trust Degree And Social Tags

With the broad popularization of Internet and hypergrowth of Information Technology,the Internet has become a significant access to obtain desired information for people and information volume in the Internet is explosively rising at an extraordinary speed.As the traditional information services,like Classified Directory and Search Engine,reveal its weakness in meeting user's individualized needs,Recommendation System emerge as the times require.Aiming to improve user experience and enhance competitiveness,personalized recommendation technologies have been broadly applied to many platforms,including the social network,E-commerce,movie and music website etc,currently.Collaborative filtering,characterized by stability and simplicity,has become one of the most widly and successfully applied personalized recommendation.But the traditional collaborative filtering still suffers from many shortages,including: 1)sparsity problem which means inaccurate measure of user similarity due to over sparse rating matrix;2)the predict result of model based on single user interest,which is widly used in traditional collaborative filtering,shows great deviation when user have multiple interests of huge span;3)the hypothesis that user have permanent interest failed to reflect the fact that user interest is dynamic in reality.Aiming to address above-mentioned problems,this thesis proposes two improved collaborative filtering algorithms based on user explicit rating and social tagging information.1)This thesis proposes a multi-interest recommendation algorithm based on user trust degree and preference of item attributes.It produces user-item attribute preference matrix based on user rates and item attributes information.Then we filter neighbors by combining user rating similarity and attribute preference similarity.Then we propose a user-level trust degree to select K closeest neighbor by merging aforementioned hybrid similarity.Finally,when comes to rating prediction,combing rate similarity and truet degree,which is proposed in this thesis based on specific users and particular items,to produce recommendation weights.2)This thesisi also propose an improved collaborative filtering based on time and social tagging.Word embedding of social tagging is generated through Word2 Vec model at first step.Then referring to the idea of Inverse Document Frequency in TF-IDF,we cluster those word embeddings to build user-tag category preference matrix.Then,we design a new time-decay function,which reflects the transfer of user interest hiddening in the dynamic tagging infromation,to represent the weight of tag category.Then we calculate user similarity for producing rate prediction,finally.A hybrid model is created by combing those two algorithms proposed in this thesis.Those collaborative filtering algorithms effectively alleviated the inaccuracy problem of recommendation caused by sparsity problem,unicity of interest model and dynamism of interest and have been proved that have better recommendation accuracy than other relevant improved algorithms.