An Explorative Study Of Recommendation Algorithm Based On Multi-criteria Ratings

In order to alleviate the pressure brought by information overload and enhance users’satisfactory, recommender systems have been widely developed to assist users in selecting therelevant part of online information or service. Multi-criteria rating recommender systems,promising to increase recommendation accuracy by utilizing multiple criteria ratinginformation, have been successfully applied in many domains. However, most techniquesrested in Multi-criteria rating recommender systems lack of considering the fact that usersemploy different preference structure for different items. To be specific, a user may employ acriteria preference to some items but a totally different one with respect to other items.In this paper, we put forward a novel approach for multi-criteria recommendation, basedon the idea of simultaneously clustering both users and items into groups which meet thatusers employ similar preference structure when making overall assessment for items withinthe same group. Having introduced the general approach of detecting users’ differentpreference structure, we propose to apply two alternative recommendation methods withineach cluster and the predictions from all clusters will be aggregated as recommendationoutput. In particular, one of the two recommendation methods is capable of eliminating thedependencies among multiple criteria so as to model user preference structure precisely andimprove the quality of recommendation. Besides, this method is expected to deal with sparsedataset and cold-start user problem. Moreover, we propose to extract multi-criteria ratingsfrom user-generated reviews and further apply them in the process of clustering as well asrecommendation, so as to evaluate its difference with those directly provided by users.To demonstrate the effectiveness of the method proposed in this paper, we collect twodifferent datasets from TripAdvisor for experiments, one is about hotel and the other is aboutrestaurant. The empirical results demonstrate that our method outperform the traditionalmulti-criteria recommendation techniques and other state-of-the-art approaches whichpostulate that users employ same preference structure for all items.