When autoencoders meet recommender systems : COFILS approach

Abstract

Collaborative Filtering to Supervised Learning (COFILS) transforms a Collaborative Filtering (CF) problem into classical Supervised Learning (SL) problem. Applying COFILS reduce data sparsity and make it possible to test a variety of SL algorithms rather than matrix decomposition methods. It main steps are: extraction, mapping and prediction. Firstly, a Singular Value Decomposition (SVD) generates a set of latent variables from a ratings matrix. Next, on the mapping phase, a new data set is generated where each sample contains a set of latent variables from an user and it rated item; and a target that corresponds the user rating for that item. Finally, on the last phase, a SL algorithm is applied. One problem of COFILS is it’s dependency on SVD, that is not able to extract non-linear features from data and it is not robust to noisy data. To address this problem, we propose switching SVD to a Stacked Denoising Autoencoder (SDA) on the first phase of COFILS. With SDA, more useful and complex representations can be learned in a Deep Network with a local denoising criterion. We test our novel technique, namely Deep Learning COFILS (DL-COFILS), on MovieLens, R3 Yahoo! Music and Movie Tweetings data sets and compare to COFILS, as a baseline, and state of the art CF techniques. Our results indicate that DL-COFILS outperforms COFILS for all the data sets and with an improvement up to 5.9%. Also, DL-COFILS achieves the best result for the MovieLens 100k data set and ranks on the top three algorithms for these data sets. Thus, we show that DL-COFILS represents an advance on COFILS methodology, improving it’s results and that is a suitable method for CF problem.