Graph Attention Network Towards Prediction and Classification of Fake Reviews in E-Commerce Application

Abstract

Nowadays digital marketplace is growing tremendously due to changes in lifestyle of peoples and rapid increase of the technologies. Especially E commerce is becoming driving growth of many businesses through its cost effective strategies. In particular, E commerce gathers user reviews to their businesses as a user interaction and user intention to enhance purchase decision, improve user experiences and to build trust on their product and services. Thus, it becomes mandatory to analyze and evaluate the user reviews on their products and services. Traditionally, opinion mining and sentiment analysis were used to analyze the sentiment associated along the reviews in form of positive and negative aspects. Despite of several advantages of those models fails to capture the fake reviews which lead to several consequences in form of consumer deception, loss of trust and unfair competitions. In order to mitigate those challenges, a new deep learning architecture is designed to detect and classify the fake reviews propagating in the e commerce platform such as amazon, bigbasket and flipcart. In this work, Convolution Neural Network along Long Short Term Memory Mechanism has been designed as deep learning model to detect the fake reviews on product reviews. Initially data preprocessing is carried out using stop word removal, stemming and weighted tokenization technique. Weighted Token in form of word vector is projected to convolution layers of the Convolution Neural Network to extract the user specific features and product specific features separately. Further those extracted user specific and product specific feature is projected to pooling layer. Pooling layer incorporates Long Short Term Memory Mechanism to identify the long dependency map among the user specific features and product specific features. Finally long dependency map is projected to fully connected layer containing softmax function to predict and classify original review and fake reviews. Experimental analysis and performance analysis of the proposed model is performed using amazon dataset. On experimental and performance analysis, it is proved that proposed model outperforms state of art approaches in classifying the original and fake reviews with accuracy of 98.7%.