Improving Predictions with User and Item Profiling

The value of Biased Matrix Factorization algorithms in recommender systems, based only on numeric ratings, has already been demonstrated. Improvements in predictions can be achieved adding more information, for example considering user generated textual reviews, although the lack of rules increases the level of difficulty in machine learning methodologies. The aim of the presented activity is to experiment the online Latent Dirichlet allocation to build user and item profiles in order to improve predictions obtained with a Biased Matrix Factorization algorithm. For the experimental analysis the Yelp data set was used, limited to the Restaurant category. Applying a 5-fold cross validation promising results were obtained in terms of Root Mean Squared Error (RMSE).