Deep learning-based approaches to handle challenges of multi-criteria recommender systems

With increasing usage of the Net, users have difficulties in accessing products/services which matching their personal tastes. Traditional recommender systems are effective methods to cope with this situation. However, unlike traditional recommender systems, users have evaluated products/services considering different criteria recently. Multi-criteria recommender systems are extensions of traditional recommender systems that allow users to evaluate considering different criteria. Accuracy of produced predictions by multi-criteria recommender systems affects users' satisfaction and sustainability of these systems. Furthermore, due to the increasing number of criteria, sparsity becomes a more prominent problem for multi-criteria recommender systems and producing predictions at a reasonable time becomes difficult due to the increasing amount of data. The success of deep learning techniques in extracting hidden, complex and nonlinear features from heterogeneous and big data has led to the frequent usage of these techniques in traditional recommender systems. In this dissertation, new approaches based on autoencoders have been proposed in order to solve main problems of multicriteria recommender systems such as accuracy, sparsity and scalability. Experimental works and analyzes performed on real data sets have shown that the proposed methods are more effective than the existing state-of-the-art multi-criteria recommendation techniques in solving the specified problems. Keywords: Multi-criteria recommender systems, Accuracy, Sparsity, Scalability, Deep learning