Optimization Study In Taxonomy-based Recommender Systems

Collaborative filtering and latent factor model is widely used to model the users’preference and recommend items to them. However, since the data sparse issue and coldstart problem are commonly encountered in real-world scenarios, latent factor modelperforms poorly at this situation. Although researchers propose social recommendationswhich integrate social network information into collaborative to remit the user cold startproblem, social recommendations still suffer the cold start problem.To solve this problem completely, Taxonomy-based recommender systems areproposed in this paper. The taxonomy information is human-labeled categories and can beused to solve the item cold start problem. The taxonomy-based matrix factorization(TBMF) model which incorporates taxonomy into matrix factorization is realized in thispaper. Generating the user-category rating matrix and linearly adding it to the users’ rating,the TBMF model can help to recommend cold start items. For social recommendation, thetaxonomy-based social trust ensemble (TBSTE) model which integrates taxonomyinformation and social network information into collaborative filtering performs well. Inthe TBSTE model, the users’ rating is a combination of their own ratings, their friends’ratings and their ratings on categories.Experiments are performed on real datasets. Experiment results on MovieLensdatasets demonstrate the TBMF model decrease the error of the basic model by3%. Itmeans that incorporating taxonomy information into matrix factorization can increase theaccuracy of recommender systems. Results on Epinions and ciao dataset demonstrate thegain of TBSTE model over basic model is6%. This means that taxonomy information canimprove the recommend performance and solve the item cold start problem.