A Deep Neural Network (DNN) Approach for Recommendation Systems

Abstract

AbstractIn different commercial platforms, Recommendation System (RS) is widely used for providing recommendations to users. In various areas, RS is utilized broadly and in E-Commerce sites, product recommendations are discovered during an active user interaction by RS. In recent decades, some key challenges are faced due to tremendous growth in user as well as products. Moreover, in RS, computation of right product and active user is a major task. User inclination and socio-demographic behavior are considered in existing works for recommending a product. In recommendation systems, one of the major algorithm used is Collaborative Filtering (CF) algorithms. This algorithm is simple as well as effective. However, further enhancement of recommendation result’s quality is limited by data sparsity and scalability of this technique. Previous technique’s problems are addressed effectively in proposed technique and user preference on balance feature analysis and products are evaluated. Therefore, proposed a model using the combination of deep learning technology and CF recommendation algorithm with three major stages, namely, preprocessing, representation of features and rating score prediction using DNN. At first, from log files, redundant and unnecessary data are removed in preprocessing module. There is an unwanted files like repeated tags, repeated similar products, removing invalid values, last visit and elapsed time. In feature representation stage, Quadric Polynomial Regression—QPR-based feature representation technique is used. The traditional matrix factorization algorithm is enhanced for obtaining accurate latent features. At last, DNN model is fed using these latent features as input data, which is a second stage of proposed model. Rating scores are predicted using this. From Amazon dataset, user data based on behavior is obtained and used in experimentation. There are 18,501 product reviews in Amazon product dataset. From Amazon web services, collected the dataset information that joins with administrative services. Based on metrics like F1-measure, Recall (R) and Precision (P) proposed Deep Neural Network (DNN) method is evaluated experimentally and highest value of those metrics are produced when compared with state-of-the art techniques like K-Nearest Neighbor (K-NN), Artificial Neural Networks (ANN).KeywordsRecommendation system (RS)Collaborative filtering (CF)Quadric polynomial regression (QPR)Deep neural network (DNN)