Fake Product Review Detection using NLP

Fake reviews distort buying decisions and marketplace trust, and they are written to sound genuine — which makes them a natural text-classification problem. Most student spam-detection projects train on toy data or reuse a black-box API, leaving the student unable to explain what the model actually learned. This project works the opposite way: it trains on the published Deceptive Opinion Spam corpus (Ott, Choi, Cardie & Hancock, 2011) — 800 fake reviews commissioned from crowd workers and 800 truthful reviews mined from TripAdvisor — and learns which linguistic cues separate them. The pipeline is deliberately transparent: tokenization, TF-IDF unigram/bigram features, and logistic regression, so every weight can be inspected and every prediction explained with its top contributing n-grams. The web demo makes the model tangible: paste any review and watch the classifier score it with a per-signal breakdown of superlatives, exclamation use, pronoun patterns, specificity and repetition.

1. The Deceptive Opinion Spam corpus is loaded: 1,600 labeled reviews (800 fake from crowdsourced writers, 800 truthful from TripAdvisor), all Chicago hotels.
2. Text is preprocessed — lowercased, punctuation-stripped, stopwords removed — then vectorized with TF-IDF over unigrams and bigrams (design target ~10,000 features).
3. The data is split 60/20/20 into train, validation and held-out test sets with stratification.
4. Logistic regression is trained on the training split while the regularization strength C is tuned with 5-fold cross-validation on the validation data.
5. The final model is evaluated exactly once on the held-out test split: accuracy, precision, recall, F1 and the confusion matrix are recorded for the report.
6. In the web demo, a pasted review goes through the identical preprocessing and vectorization, and the trained model returns P(fake) with per-signal linguistic scores for the UI breakdown.

Tech stack:

• Python 3, scikit-learn (TF-IDF, logistic regression, cross-validation)
• pandas, NumPy, Matplotlib (analysis & plots)
• Jupyter Notebook (training & evaluation)
• HTML5 + CSS3 + vanilla JavaScript (interactive analyzer demo)
• Deceptive Opinion Spam corpus (Ott et al., 2011)

Dataset & model details

• Dataset: Deceptive Opinion Spam corpus (Ott, Choi, Cardie & Hancock, ACL 2011) — 1,600 hotel reviews: 800 fake (written by Amazon Mechanical Turk workers) and 800 truthful (from TripAdvisor), all for Chicago hotels, balanced classes.
• Task: Binary text classification; input = raw review text, output = P(fake) with per-signal linguistic scores.
• Model: TF-IDF vectorizer (unigrams + bigrams, design target ~10k features) feeding L2-regularized logistic regression; C tuned by 5-fold cross-validation.
• Metrics: Test accuracy ~89% (design target, matching published bigram-model benchmarks on this corpus — the built-to-order run's measured figure is reported after training), plus precision, recall, F1 and confusion matrix on the held-out 20% split. No accuracy is claimed as measured until the training run is executed for the order.

• [Deceptive Opinion Spam corpus pipeline] Loads the real 1,600-review corpus (800 fake / 800 truthful Chicago hotel reviews), with documented train/validation/test splits and class-balance handling.
• [TF-IDF + logistic regression model] Transparent, inspectable classifier over unigram and bigram features — every prediction is explainable through its top-weighted n-grams, ideal for viva questions.
• [Interactive review analyzer demo] Paste any product or hotel review; the demo returns a fake/genuine verdict with confidence and a six-signal linguistic breakdown.
• [Top-signal inspection] The notebook lists the highest-weighted fake-indicative and genuine-indicative n-grams, turning "what did the model learn?" into a viva-ready discussion.
• [5-fold cross-validation] Regularization strength tuned on validation folds, with the full procedure documented for reproducibility.
• [Held-out evaluation] Accuracy, precision, recall and F1 reported once on a 20% test split that the model never saw during tuning.
• [Sample review gallery] Pre-loaded suspicious and genuine samples so the demo is impressive from the first click, plus a session history table of analyzed reviews.

• Complete training & evaluation Jupyter notebook
• Trained TF-IDF vectorizer + logistic-regression model files with inference code
• Interactive review-analyzer web demo with sample reviews and signal breakdown
• Top-weighted n-gram lists (fake-indicative vs genuine-indicative) with interpretation notes
• Project report PDF (background, NLP theory, corpus description, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (TF-IDF, n-grams, regularization, precision vs recall, overfitting)

• Trained on hotel reviews — performance on other domains (electronics, restaurants) is expected to drop without retraining, stated honestly in the report.
• The model reads linguistic style, not facts: a truthful but flowery review can score as suspicious, and a careful liar can evade it.
• ~89% is a design target matching published benchmarks, not a measured claim — the report documents the actual achieved figure after the training run.
• Very short reviews (under ~20 words) carry too little signal for a reliable verdict; the demo warns on these.
• This is a teaching build for fake-review linguistics, not a production moderation system.