Hybrid Model Rating Prediction with Linked Open Data for Recommender Systems

Abstract

We detail the solution of team uniandes1 to the ESWC 2014 Linked Open Data-enabled Recommender Systems Challenge Task 1 (rating prediction on a cold start situation). In these situations, there are few ratings per item and user and thus collaborative filtering techniques may not be suitable. In order to be able to use a content-based solution, linked-open data from DBPedia was used to obtain a set of descriptive features for each item. We compare the performance (measured as RMSE) of three models on this cold-start situation: content-based (using min-count sketches), collaborative filtering (SVD++) and rule-based switched hybrid models. Experimental results show that the hybrid system outperforms each of the models that compose it. Since features taken from DBPedia were sparse, we clustered items in order to reduce the dimensionality of the item and user profiles.