Employment Fraud Detection System: Safeguarding Job Search Security with NLP and Machine Learning

Project Core: An employment fraud detection system based on NLP and machine learning, designed to help job seekers identify fake recruitment information. Through TF-IDF feature extraction, Logistic Regression/XGBoost and other models, combined with an interpretability mechanism, it has been deployed as a Streamlit online application and open-sourced on GitHub (author: nikhilasds25-bit, release date: 2026-06-13).

In the era of digital recruitment, fake job advertisements are rampant: criminals deceive job seekers with methods such as high salaries, advance fees, and "guaranteed employment", causing millions of economic losses and time waste every year. Developer Nikhil A S built this system to reduce the risk of job seekers being scammed through automated analysis.

Dataset: Uses the Fake Job Postings Dataset (17,880 records), with class imbalance (17,014 real/95.16%, 866 fake/4.84%). Feature Engineering: Merges text fields such as job title and description into feature vectors, and introduces structured features (company logo existence, screening questions, remote work indicator, etc.) to improve performance.

Technical Process: Convert text to numerical vectors using TF-IDF. Model Iterations:

• Version 1: Logistic Regression (accuracy 97%, fraud recall 88%)
• Version 2: XGBoost (accuracy 98%, recall 63%)
• Version 3.1: XGBoost + structured features (recall 69%)
• Version 3.2: Logistic Regression + structured features (recall 90%, suitable for scenarios sensitive to missed detections).

Version 4 introduces a trust score (0-100 points), with dimensions including: company logo completeness, screening strictness, work mode (remote work has high risk), and model confidence. The risk explanation system generates readable reasons (e.g., "missing company logo") to improve transparency.

Deployment: Implemented as an online web application via the Streamlit framework, supporting interactive input, real-time analysis, confidence visualization, trust score display, and risk explanation. Application Scenarios: Pre-submission screening for job seekers, auxiliary review for recruitment platforms, HR risk early warning.

Technical Highlights:

1. Class Imbalance Handling: Uses sampling strategies, focusing on fraud recall rate;
2. Feature Insights: Fake jobs have higher no-logo rate (7.4% vs real 4.1%), lower no-screening-question rate (28.8% vs real 50.2%), lower remote work rate (32.7% vs real 81.9%);
3. Interpretability: Allows users to understand the basis for judgments, avoiding reliance on black boxes.

Future Directions: At the model level, introduce pre-trained language models (DistilBERT, etc.) and ensemble learning; at the engineering level, develop REST API, multilingual support, real-time portal integration, and company verification processes; deepen explainable AI. Social Value: Protect job seekers' rights and interests, purify the recruitment ecosystem, promote technology inclusion, and demonstrate the application of NLP in social governance.