Virtual Booth

Online shopping has become increasingly popular due to its ease and variety of choices, encouraging users to leave online reviews. However, the growing presence of fake reviews poses a threat to consumer trust and can mislead potential buyers. Many existing models for detecting fake reviews focus on the entire text and often miss subtle indicators such as overly emotional language, repeated phrases, or vague content. Another common issue is the imbalance in review datasets, where genuine reviews significantly outnumber fake ones, resulting in models that are less effective at spotting deceptive content. To address these challenges, this study proposes a hybrid approach that combines Implicit Aspect-Based Sentiment Analysis (ABSA) with techniques for handling imbalanced data. Implicit aspects are extracted, and their sentiments are analysed using Bidirectional Encoder Representations from Transformers (BERT), and the Synthetic Minority Over-sampling Technique (SMOTE) is used to balance the dataset. Rule-based indicators help flag suspicious reviews, and a Support Vector Machine (SVM) classifier is used for final classification. Its performance was compared with an SVM baseline model and a BERT + Rule-based + SVM without the SMOTE approach. The constructed hybrid approach achieved strong results, particularly with 100% recall, meaning it successfully identified all fake reviews. Although precision was slightly lower, the high recall is valuable in contexts where missing fake reviews can have serious consequences. This study supports better decision-making in e-commerce and recommends future work on expanding datasets, refining resampling methods, and applying the approach in fields like healthcare or education.