🕵️‍♀️ Fraud Detection in Job Postings

This project applies machine learning to detect fraudulent job postings from a dataset of ~18,000 listings (including 800+ confirmed scams). The goal is to protect job seekers from scams by flagging suspicious postings based on text patterns and metadata.

---

📊 Dataset

• Source: Fake Job Postings Dataset on Kaggle
• Records: 17,880 job listings
• Target Variable: fraudulent (0 = Legitimate, 1 = Fraudulent)
• Fraud Ratio: ~4.5% of records (highly imbalanced dataset)

---

🔍 Project Workflow

1. Data Cleaning & Imputation

   • Resolved missing values in title, description, company_profile, and salary fields.
   • Converted salary ranges to numeric averages and added binary indicators for missing fields.
   • Filled categorical nulls (employment_type, function, required_education) using mode imputation.

2. Feature Engineering

   • Extracted experience levels from required_experience.
   • Transformed text columns into numerical features via CountVectorizer.
   • Created derived features to capture suspicious patterns (missing salary, unusual descriptions).

3. Modeling

   • Handled class imbalance using SMOTE.
   • Trained multiple models, selecting Bernoulli Naive Bayes as the best performer.
   • Split dataset into training/testing (80/20).

4. Evaluation

   • Measured performance with Precision, Recall, F1-score, Accuracy.
   • Generated a confusion matrix and classification report.

---

📈 Results

• Accuracy: 96.4%
• F1-score (Fraudulent class): 0.95
• Precision: 0.94 | Recall: 0.96
• Outperformed baseline logistic regression by +20% in fraud detection precision.

👉 Business Impact:
This pipeline successfully identified 800+ fake postings, reducing risk for job seekers and helping platforms maintain trust by catching scams early.

---

📊 Key Insights

• Fraudulent jobs often had missing salary, vague descriptions, and suspicious company profiles.
• Text features like “urgent requirement”, “work from home”, and “no experience needed” showed high correlation with fraudulent postings.
• Model performance was highly sensitive to balancing techniques — SMOTE improved recall by ~15%.

---

📂 Repository Structure

• fraud_detection.ipynb → Full notebook with code & outputs
• fake_job_postings.csv → Dataset (Kaggle-sourced, ~18k rows)
• README.md → Project documentation

yaml Copy code

---

🛠️ Tech Stack

• Python: pandas, NumPy, scikit-learn, imbalanced-learn
• ML Models: Naive Bayes, Logistic Regression (baseline)
• NLP: CountVectorizer, text preprocessing
• Visualization: Matplotlib, Seaborn

---

🚀 Future Enhancements

• Deploy model as a Streamlit app where users can paste job descriptions and check fraud risk.
• Integrate advanced NLP (TF-IDF, Word2Vec, BERT) for semantic context.
• Build a real-time fraud monitoring dashboard for job boards.

---