A New Modeling for Item Ratings Using Landmarks
Gustavo Rodrigues Lima, Carlos E. Mello, Geraldo Zimbrão · 2018
Collaborative Filtering CF has been a widely used approach for personalized recommendations. In this context, model-based CF algorithms have been studied extensively in the literature and have shown higher accuracy in rating prediction than memory-based ones. The major approach regarding model- based CF is Matrix Factorization (MF). It uses the item ratings given by users to predict unknown ratings, which are posteriorly used in recommendations. Usually, MF learns a data model through optimization algorithms by requiring too much time to process. To overcome this issue, we propose a novel modeling for item ratings that allows one to apply Supervised Learning (SL) algorithms to predict unknown ratings. This modeling consists in computing item similarities to a preselected set of items, namely landmarks. Then, we build a model for each user using his/her ratings as labels and the corresponding vector of item similarities to landmarks as samples. In this work, it was applied Support Vector Regression to learn the data model. We compared our proposal against 5 state-of-the-art CF techniques in 6 different databases. The results show the proposal is able to reduce computational time and still keep competitive accuracy compared to the experimented CF algorithms.