Benchmarking Supervised Learning, Contextual Transformers, and Explainability Layers for Online Recruitment Fraud Detection

Abstract

Digital hiring platforms make applicant sourcing fast and scalable, but they also give bad actors an easy way to distribute fraudulent job postings. Detecting Online Recruitment Fraud (ORF) automatically remains difficult because scam texts mimic authentic corporate ads, and natural datasets suffer from severe class skew (scams make up less than 5% of postings in the wild). In this study, we survey recent detection techniques and benchmark several model families—from traditional shallow baselines and recurrent nets to bidirectional Transformers—against the 17,883-record Employment Scam Aegean Dataset (EMSCAD). We evaluate Word2Vec semantic projections, address minority-class starvation with SMOTE oversampling, and assess model interpretability via SHAP and LIME. Regularized XGBoost paired with SMOTE and dense Word2Vec features yielded 99.44% accuracy, a 0.99 F1-score, and an AUC of 0.9999. Pretrained Transformer encoders (RoBERTa and BERT) also performed strongly (98.81% and 98.67% accuracy, respectively) while better tolerating linguistic variations. Because deep neural nets and boosted tree ensembles do not provide transparent reasoning out of the box, we couple these models with local attribution plots to build an auditable inspection pipeline for real-world hiring platforms. Keywords— Online Recruitment Fraud, Transformer Architectures, BERT, XGBoost, Explainable AI, SMOTE, EMSCAD, Hiring Security.