CUSTOMER REVIEWS ANALYSIS WITH DEEP NEURAL NETWORKS FOR E-COMMERCE RECOMMENDER SYSTEMS

Abstract

The first part of this thesis systematically reviews the trend of researches conducted from 2011 to 2018 in terms of challenges and problems regarding developing a recommendation system, areas of application, proposed methodologies, evaluations criteria used to assess the performance and limitations and drawbacks that require investigation and improvements. The study provides an overview for those who are interested in this field to understand the current and the future research opportunities. The second part of this thesis proposes a new methodology to consider customer reviews in recommender systems. An essential prerequisite of an effective recommender system is providing helpful information regarding users and items to generate high-quality recommendations. Customer reviews are a rich source of information that can offer insights into the recommender systems. However, dealing with the customer feedback in text format, as unstructured data, is challenging. Our research includes extraction of the features from customer reviews and use them for similarity evaluation of the users to generate the recommendations. To do so, we have developed a glossary of features for each product category using Latent Dirichlet Allocation. We then employed a deep neural network to extract deep features from the users-attributes matrix to deal with sparsity, ambiguity, and redundancy. Furthermore, we then applied matrix factorization as the collaborative filtering method to provide recommendations. The experimental results using Amazon dataset demonstrate that our methodology improves the performance of the recommender system by incorporating information from reviews when compared to the baselines.