Performance Analysis of Deep Learning Architectures for Recommendation Systems

Abstract

Recommendation systems play an important role in the field of e-commerce applications since they provide suggestions to each and every customer based on the reviews and ratings given by the customers. These reviews and ratings allow customers to share their opinions and experiences about products they purchase. This enables companies to market to more people of a similar demographic and influence more purchases. Deep learning techniques with different neural network architectures can be applied to the recommendation systems to identify the different patterns and behaviours of the customers in e-commerce applications. The main aim of this paper is to study the effect of combining deep learning neural architectures and collaborative filtering to provide an effective recommendation system. A comparative study of natural language processing techniques is analysed using three different Recurrent Neural Network (RNN) models that convert reviews to ratings. The RNNs that are included are Long Short Term Memory (LSTM), Gated Recurrent unit (GRU) and lastly, a multilayer RNN that includes LSTM stacked with GRU to test the possible advantages of a deeper architecture. A Neighbourhood based Collaborative Filter Recommendation System is developed that gives recommendations to users based on item-item similarities. The performance of the three models is analysed to find the best model to perform Review Rating prediction in order to enhance the accuracy of the Recommendation system.