Abstract
Deep learning based neural networks have been attracting significant interest lately, due to their success in complex automatic recognition tasks in many artificial intelligence areas such as language recognition ,computer vision and expert systems. However, in recommendation systems, they have not been exploited fully, and most of the systems rely on traditional collaborative filtering with matrix factorization approaches. In this paper we suggest a novel approach based on deep learning-based augmentation of the collaborative filtering approach for predicting user ratings for different types of media collections in online databases and libraries, including movies, music and book collections. Though there have been few approaches proposed based on deep learning for recommender systems, they were mainly used for modelling additional complementary information available in terms of images, text or acoustic information from different items in the collections. The main component, the interaction between the item and the user was still modelled with traditional matrix factorization approaches, which uses the dot product to extract the latent features between users and items. By augmenting the latent features extracted with deep learning based neural network models, it is possible to capture the hidden interactions and can lead to better recommender system performance even with sparse and unbalanced datasets. The experimental evaluation of the proposed approach based on four different publicly available datasets involving movies, music and book collections, show promising results, in terms of different performance metrics including accuracy and mean absolute error. Experiments on four datasets show the effectiveness of our proposed frameworks.