Fake Review Detection for Online Electronics Marketing using Hybrid Deep Neural Network Model

Abstract

Internet has transformed the domestic market into an online business, and customer feedback has made a difference in determining a company's e-commerce revenue. The majority of consumers read reviews given by previous reviewers before making any purchase decision on products or services. Trustworthiness of online reviews is critical for organizations; it has an indirect or direct impact on industry profitability. Because of this impact, several companies use spammers to post fake reviews and also influence customer decisions. Fake review identification is a challenging and time-consuming process that requires a lot of resources in the area of electronics consumers.. And most of the prior research was focused on detecting reviews in the domains of hotels, health, mobile apps, and restaurants. However, existing fake review detection studies that use machine learning and deep learning techniques are ineffective due to issues such as low detection performance, unbalanced datasets, and small attributes of detection review used to build models. And, it ignores the context of words in a review and focuses on extracting traditional linguistic units from reviews. The attribute of review used in earlier studies was insufficient to identify fake review content, and the focus of review authentication was on the reviews' individual word meanings. This study proposed a fake review detection using a hybrid deep neural network model that combines CNN and BILSTM with word embedding. For experiments, this study used publicly accessible labeled datasets from Amazon electronic consumers. The study was carried out using sentiment analysis as the feature used to calculate the polarity and subjectivity of the reviews. The proposed hybrid CNN-BILSTM model extracts features from the input review contents using CNN layers and sequence learning using BILSTM layers. The study used the functional Keras API, which concatenated review input types such as behavioral and contextual features together. An effective feature extraction technique and more attributes were used to increase detection performance. In this study, we compared different deep neural network-based algorithms to find out which is the most important for fake review detector model development, such as CNN, BILSTM, MLP, CNN-MLP, CNN-GRU, and CNN-LSTM. Finally, based on our experiment, we found that the CNN-BiLSTM model performed the best for the proposed fake review detection, with an accuracy of 96.70%, which is 3.40% higher than the previous study and an f1-score of 98%.