Detection of Fake Reviews in Yelp Dataset Using Machine Learning and Chain Classifier Approach

Over the past few years, e-commerce has led to a significant shift in business activities from traditional methods to online platforms. Nowadays, consumers heavily rely on online reviews to guide their purchasing decisions, prompting businesses to adapt to this new reality. However, fake reviews pose a critical challenge in online reviews. Fake reviews can have serious consequences, including misleading customers and damaging the reputation of organizations. To tackle this issue, various approaches, such as natural language processing, machine learning, and sentiment analysis, have been proposed as potential solutions for detecting fake reviews. These strategies typically involve analyzing the content of reviews along with associated metadata, such as the language used, review timing, and ratings. However, differentiating between fake and genuine reviews can be challenging, as fake reviewers often employ tactics to make their reviews appear more legitimate. Despite the complexities involved, significant progress has been made in developing effective strategies for detecting fake reviews. These techniques play a crucial role in ensuring that consumers can make informed decisions based on trustworthy information while safeguarding online review systems' integrity. The main focus of this paper is to combine textual elements with other related behavioral parameters, which leads to a higher rate of perception and detection compared to other existing methods. By incorporating new behavioral variables, the proposed model enhances the accuracy of detecting fake reviews. The Elmo Model is utilized for encoding result vectors and reducing computational overhead, while the VADER model helps to determine the polarity of review text, enabling individual reviews to be evaluated. To classify real-time reviews in the Yelp dataset, a stack model is constructed using the Multinomial Naive Bayes method (MNA) and the Gradient Boosting Classifier. Compared to similar studies, the proposed model demonstrates exceptional accuracy, achieving an AUC of 82% and overall accuracy of around 98%. Overall, the integration of textual components with behavioral parameters in the proposed model offers a promising approach to effectively detect fake reviews and improve the authenticity of online review systems.