Live Interactive Demo: Try HyNieSVM Online
HyNieSVM is a machine learning and natural language processing system designed to detect deceptive and computer-generated online reviews across e-commerce and service platforms.
The system uses a late-fusion ensemble approach combining linguistic pattern recognition with knowledge-aware entity verification to distinguish between authentic customer feedback and generated/fake reviews.
HyNieSVM integrates three core components:
- DistilRoBERTa: Extracts dense contextual and stylometric features with high inference throughput.
- ERNIE 2.0 / 3.0: Captures factual and entity-level consistency to detect hallucinated review attributes.
- Late-Fusion Concatenation: Combines both encoder outputs into a unified 1536-dimensional feature representation.
- Linear Support Vector Machine (LinearSVC): Provides fast, convex boundary classification with O(n) scalability, calibrated for probabilistic outputs.
- GWO-PSO Optimizer: Uses a hybrid metaheuristic search to identify optimal regularization parameters for the SVM classifier.
Evaluated on the Kaggle Fake Reviews Dataset using an 80/20 stratified split:
| Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| K-Nearest Neighbour | 53.36% | 51.70% | 92.50% | 66.33% |
| SiEBERT (Sentiment-Only) | 64.73% | 64.38% | 64.87% | 64.62% |
| TF-IDF + XGBoost | 82.65% | 85.24% | 78.69% | 81.83% |
| Naive Bayes | 83.78% | 80.91% | 88.12% | 84.36% |
| SVM + RoBERTa | 92.94% | 92.97% | 92.80% | 92.88% |
| HyNieSVM (Proposed System) | 94.72% | 94.63% | 94.75% | 94.69% |
- Clone the repository:
git clone [https://github.com/](https://github.com/)<your-username>/HyNieSVM.git cd HyNieSVM