neOptimizing Deep Learning Model CNN-BiLSTM for Fake Product Review Classification

The increase in e-commerce users makes product reviews one of the triggers for consumers to make purchasing decisions. Helpful product reviews is critical for enhancing user experience and trust in e-commerce platforms. This study presents a review-centric approach to optimizing deep learning models for classifying product reviews as “trusted,” “fake,” or “non-review.” We collected a dataset of 2324 skincare product reviews from popular review platforms femaledaily.com, through web crawling. We implemented a hybrid CNN-BiLSTM model. Hyperparameters like epoch, learning rate, batch size, and dropout rate were fine-tuned to enhance model performance. The models were evaluated and analyzed using metrics including accuracy metric, then precision, recall, F1-score, and use confusion matrix. The outcomes of our experiments reveal that the hybrid model CNN-BiLSTM model outperformed other architectures, achieving the highest accuracy of 94%. These findings highlight the effectiveness of the CNN-BiLSTM for the accurate classification of helpful skincare product reviews and offer insights into deep learning optimization strategies for fake review detection.