Fake Review Detection on Indonesian E-Commerce Platforms Using a Hybrid CNN-BiLSTM-Attention Architecture Optimized by Differential Evolution

Abstract

Fake reviews on e-commerce platforms threaten consumer trust and harm local micro, small, and medium enterprises (MSMEs). This study develops a fake review detection system using a hybrid deep learning architecture—Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and an Attention mechanism—whose hyperparameters are optimized by Differential Evolution (DE). A two-phase design is proposed: Phase 1 validates the architecture on an English benchmark (Kaggle Fake Reviews, 40,432 reviews), and Phase 2 generalizes it to an Indonesian e-commerce corpus (38,825 cleaned reviews from Tokopedia and Shopee) labeled automatically via Latent Dirichlet Allocation (LDA). Because no Indonesian corpus records review authorship, the validity of the induced labels is treated as a research question rather than an assumption: the topic-to-class mapping is anchored to blind annotation by two independent annotators, and agreement is quantified with Cohen's Kappa on a held-out half of the annotated sample. The proposed model was compared against four baselines (Standalone LSTM, CNN+PSO, BiLSTM+GA, and Standalone DE) across six metrics over independent runs, with One-Way ANOVA and Tukey HSD post-hoc tests. The proposed model achieved competitive accuracy with no statistically significant difference from the strongest deep-learning baselines. Accuracy on the Indonesian phase should be read with caution, since labels induced from word distributions are predictable from the same textual regularities the classifier learns, while Standalone DE without a hybrid architecture performed significantly worse. The main contributions are an openly documented labeling and validation procedure for Indonesian OR/CG data, an equal-budget comparison of DE against PSO and GA, and a cross-dataset generalization analysis situated in the MSME context.