Multi Feature Space LDA for Tag Recommendation in Cold Start Problem

Abstract

Social tagging is a popular way of organizing, sharing and browsing information between groups and individuals. Users of social bookmarking sites annotate resources with keywords called tags. These tags become part of resource and user pro les. Tags are useful in many applications like information seeking, information representation, and search etc. Tags are also helpful in developing user friendly interfaces like tag clouds and faceted search. Even though information is increasing day by day, still in the case of sharing new resources (the cold start problem), problems occur during the process of tagging because the sharing platform does not have any existing information about the newly added resource. In this thesis we propose a novel model MFS-LDA in tag recommendation for the cold start problem. MFS-LDA does not ignore relevant tags by separating the contents, the titles in the model. In addition, due to separation of feature spaces MFS-LDA can disambiguate the context of the resource. MFS-LDA captures human classi cation by including tags feature space in the model. Moreover, MFS-LDA introduces dependency among topics of di erent feature spaces. Dependency helps in recommending subjective tags by considering opinion from each feature space in topic assignment. In addition, dependency also helps in the disambiguation process. Dependency in MFS-LDA helps in recommending blend of context by recommending tags from each feature space. v We evaluate MFS-LDA on various baselines. With the use of dependency, MFSLDA recommends more subjective tags instead of generic tags. We evaluate our model on a large dataset consisting of \20,578" unique resources. MFS-LDA shows a signi cant improvement over di erent baselines. In the end, MFS-LDA outperforms di erent baselines even if a feature space is missing.