Fake Review Checker
Paste a product review and three trained models (CNN, BERT, XLNet) decide if it looks fake. Everything runs in your browser.
Why this result?
What pushed the score
Word influence
red words pushed towards FAKE · green words pushed towards GENUINE · darker = stronger
Explanation method: the CNN re-scores the review with each word removed and with each review feature reset to the training average; the change in fake probability is that item's influence. BERT and XLNet may weigh things differently. The comparisons describe tendencies in the training data, not rules: no single feature such as length proves a review is fake or genuine.
| Model | Accuracy | F1 (fake) | ROC-AUC |
|---|---|---|---|
| CNN | 0.930 | 0.569 | 0.885 |
| BERT | 0.930 | 0.580 | 0.884 |
| XLNet | 0.932 | 0.583 | 0.887 |
Trained on 291,762 Amazon reviews (Cell Phones, Digital Music, Electronics, Industrial & Scientific, Software). "Fake" follows the project's labelling rule: reviews by reviewers whose non-verified-purchase ratio is above 0.5. The first visit downloads about 250 MB of models; later visits use the browser cache. Final-year project, Department of Information Technology, Government College of Technology · source code