Recommender System using Deep Neural Networks

Abstract

Recommender system helps users find items of their interest, by filtering and suggesting relevant items. Collaborative filtering (CF) is one of the key methods for personalized recommendation that models the users’ preferences for items on the basis of their past interactions (such as clicks, views, ratings, reviews etc.). CF is prone to both the cold-start and sparsity problems due to insufficient interaction records. While the most well-known CF technique is Matrix factorization (MF), deep neural networks have been recently employed in CF for modeling user-item interactions. However, most of these techniques consider only the implicit/ explicit feedback for recommendation and do not give any preferences to the auxiliary information that will help in better recommendation even under sparse conditions. This thesis proposes a deep learning based model named Neural Hybrid Filtering (NHF) that utilizes auxiliary information for addressing the cold-start and sparsity problems. NHF incorporates the contextual features of items and learns two different interaction vectors by applying linear and non-linear kernels, and then merge them to calculate prediction scores. The linear kernel performs element-wise multiplication of the user and item vectors to generate the interaction vector. The non-linear kernel combines the user and item vectors through concatenation and feeds into multi-layer perceptron to get the interaction vector. Based on NHF, this thesis also proposes the Convolutional Neural Hybrid Filtering Model (CNHF). Instead of using simple concatenation, CNHF uses two-dimensional interaction map in the non-linear kernel to incorporate the user and item feature information. The interaction map is then passed into multi-layer convolutional neural network (CNN) for learning complex user-item interactions. Experiments on two publicly available datasets and quantitative comparisons with other existing models shows the superiority of our proposed models.