Fake Job Posting Detection using NLP

Employment scams have industrialized: fake recruiters post convincing job ads, move victims to WhatsApp or Telegram, and extract "registration fees" before vanishing. Job boards struggle to review every posting manually, and keyword blocklists are trivially evaded by rewording. This project treats scam detection as what it is — a text-classification problem — and solves it with a fine-tuned transformer. Trained on EMSCAD's 17,880 labeled postings, the model reads the full posting text and outputs a fraud probability, while token-attribution highlights show which phrases pushed the decision. The demo makes this tangible: paste any posting, get a risk score, and see the scam markers lit up in the text. The build is careful about its limits — it is a screening aid for platforms and job seekers, not a legal verdict on any employer.

1. EMSCAD postings are cleaned: HTML stripped, fields (title, description, requirements, company profile) concatenated into one document per posting.
2. Text is tokenized with the BERT tokenizer (≤512 tokens, truncation/padding) and split with stratification into 80/10/10 train/validation/test.
3. A BERT-base (or DistilBERT) classifier head is fine-tuned with class-weighted cross-entropy; early stopping on validation PR-AUC.
4. The held-out test set is scored: fraud-class F1, PR-AUC and false-positive rate are computed for the report.
5. In the demo, pasted text runs through the same pipeline; the model returns a fraud probability and token attributions highlight scam phrases.
6. Scores are banded with guidance text so job seekers know what to do with each verdict.

Tech stack:

• Python 3, Hugging Face Transformers (BERT/DistilBERT)
• PyTorch, scikit-learn
• pandas, NumPy
• HTML/CSS/JavaScript analysis demo
• EMSCAD dataset (Kaggle)

Dataset & model details

• Dataset: EMSCAD (Employment Scam Aegean Dataset) — 17,880 real job postings with binary fraudulent/legitimate labels (866 fraudulent, 17,014 legitimate); fields include title, description, requirements, company profile, employment type, location.
• Task: Binary text classification; input = concatenated posting text, output = P(fraudulent).
• Model: BERT-base fine-tuned (DistilBERT option for the lightweight demo); class-weighted loss, stratified 80/10/10 split.
• Metrics: Fraud-class F1 0.91, false-positive rate 0.7% (design targets for the built-to-order fine-tuning run); PR-AUC as the primary metric. No measured figures are claimed before the run.

• [BERT fraud classifier] Fine-tuned transformer on EMSCAD; input = full posting text (≤512 tokens), output = fraud probability.
• [Scam-phrase highlighting] Token-attribution surfaces the exact phrases driving each decision, rendered as highlights in the analyzed posting.
• [Risk scoring with bands] Postings scored 0–100 and banded (legitimate / suspicious / likely fraudulent) with plain-language guidance per band.
• [Scam-anatomy analysis] Dashboard views showing which phrase families separate fraud from legitimate postings across the dataset.
• [Imbalance-aware training] Class-weighted loss and stratified splits handle the 4.8% fraud rate; PR-AUC is the primary metric, not raw accuracy.
• [Legitimate-vs-fraud examples] Curated example postings demonstrate both extremes for viva demonstrations.
• [Explainability report] Per-decision phrase attributions plus corpus-level signal rankings, documented for the report.

• Fine-tuning notebook (preprocessing → training → evaluation)
• Fine-tuned model weights and tokenizer config
• Analysis web demo with phrase highlighting
• Corpus-level scam-pattern analysis plots
• Project report PDF (background, transformer theory, scam typology, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (BERT, attention, fine-tuning, class imbalance, PR-AUC)