ENHANCE FAKE REVIEW DETECTION: A HYBRID APPROACH OF IMPLICIT ABSA AND IMBALANCED DATASET HANDLING

Online shopping’s convenience has increased reliance on reviews, but fake reviews undermine trust in e-commerce. Many detection models analyze full review text, while overlooking subtle cues like lack of specificity, repetitive wording, and exaggerated sentiments expressed implicitly. Additionally, where genuine reviews far outnumber fake ones, dataset imbalance leads to biased machine learning models with reduced reliability. To address these challenges, this study introduces a hybrid approach integrating Implicit Aspect-Based Sentiment Analysis (ABSA) using Bidirectional Encoder Representations from Transformers (BERT) for implicit aspect extraction and Sentiment Analysis with the Synthetic Minority Over-sampling Technique (SMOTE) for handling imbalanced data. Rule-based indicators identify fake reviews, while a Support Vector Machine (SVM) with k-fold cross-validation evaluates performance. The dataset, sourced from Kaggle’s Amazon Reviews, contains 2,852 reviews across four product categories: foods, home care, personal care, and refreshments. Before SMOTE, the average k-fold recall was 78%. After SMOTE, it rose to 95%, enhancing the approach’s ability to detect most fake reviews despite some false positives. The final result of the constructed hybrid approach achieved 96% accuracy, 60% precision, 100% recall, and a 75% F1 score. We evaluate the performance against two comparative feature approaches, which are (i) an SVM baseline and (ii) a BERT + Rule-based + SVM without SMOTE. High recall ensures effective fake review detection, though lower precision results in some false positives. It concludes that this study enhances trust in online reviews and supports informed purchasing decisions. Future research should expand labelled datasets and explore alternative techniques like Edited Nearest Neighbors to refine the precision-recall trade-off.