High performance fake review detection using pretrained DeBERTa optimized with Monarch Butterfly paradigm

Abstract In this era of internet, e-commerce has grown tremendously and the customers are increasingly relying on reviews for product information. As these reviews influence the purchasing ability of the future customer, it can give a positive or negative impact on the businesses. The effectiveness of online reviews is compromised by fake reviews that provide false information about the product. Fake reviews can not only impact the reputation of the businesses but also involve financial losses. Thus, detection of fake reviews is essential to solve the problem for maintaining the integrity of online reviews. Existing Machine learning models often struggle with deep contextual understanding. Scaling machine learning models while maintaining accuracy and efficiency becomes increasingly challenging as the volume of online reviews continues to grow. Hence, this research work introduces a novel MBO-DeBERTa, a deep neural network with Monarch Butterfly Optimizer. The proposed model improves the capacity to differentiate between overlapping characteristics of fake and authentic reviews. MBO-DeBERTa attained a classification accuracy of 98% for detecting the fake reviews. The proposed framework is tested on three different datasets such as Amazon, Fake Review and Deceptive Opinion Spam containing 21000,40000 and 1600 reviews respectively which are publicly available in Kaggle. The proposed model also detects adversarial attacks using the Fast Gradient Sign Method (FGSM) and thereby evaluating its resistance to such attacks and noise. The proposed model was also tested on the unseen data of Myntra and Amazon verified customer reviews and our model works efficiently for real world data. Thus the results show that the suggested model outperforms the current models showing increased accuracy, precision, recall, F1 score and reduced loss rate.