Detecting Fraud in Moroccan E-Commerce Platforms Using an Explainable Deep Learning Model for Arabic Fake Reviews Classification

Fake reviews are a significant challenge in e-commerce, affecting consumer trust and decision making. This study aims to help consumers defend against fake reviews on Moroccan e-commerce platforms by introducing the Moroccan Dialectal Arabic Fake Reviews (MDAFR) dataset, consisting of 300 fake and 300 real reviews, primarily written in Moroccan dialectal Arabic. Furthermore, this study compares the performance of three deep learning models, namely LSTM, 1D CNN, and GRU, with two BERT-based models, MARBERTv2, DarijaBERT, and ArabicBERT. In addition, this research evaluates the decisions of the proposed model using the explainable methods LIME and SHAP. Using the proposed dataset, this study developed an explainable GRU model capable of detecting fake reviews in Moroccan e-commerce platforms with an average accuracy of 86.5% and an average F1-score of 86.58% across seven folds. The results showed that some keywords can increase the probability of classifying a review as fake. By examining the predictions of our model, this study identified the most common words used in fake reviews. These keywords will help consumers better identify fake reviews, thereby increasing their awareness and ability to detect misleading information. This research addresses the necessity for a fake review detection model that can capture the distinctive characteristics and cultural nuances of each dialect across Arabic speaking countries.