Automatic Fraud Detection on Employment Scam Datasets: A Machine Learning Study on the Impact of Company Information

Online recruitment has made job sourcing efficient, but fraudulent job advertisements have emerged as a serious threat that compromises the credibility of recruitment platforms, endangers the privacy of job seekers, and damages the reputation of legitimate organizations. Manual moderation of these postings is infeasible at scale. This paper presents a systematic machine learning study for automatic detection of employment scams on the Employment Scam Aegean Dataset (EMSCAD), which contains 17,880 real-world job postings (17,014 genuine and 866 fraudulent) collected between 2010 and 2014. We evaluate five conventional classifiers (Logistic Regression, Decision Tree, K-Nearest Neighbors, Support Vector Machine, Naive Bayes) and four ensemble methods (Random Forest, AdaBoost, Gradient Boosting, XGBoost) on a balanced 1,600-sample subset using TF-IDF text features and one-hot encoded structural attributes. Our central research question is whether company-information features—specifically the presence of a company logo and a company profile—improve fraud detection. Through controlled experiments comparing models trained with company information (CI) against those trained without it (WCI), and validated with 10-fold cross-validation, we find that company information improves performance for nearly all classifiers. The best configuration, Random Forest with company information, achieves 93% accuracy and a 0.93 F-measure, slightly exceeding the prior result on this dataset reported by Vidros et al. (2017). Feature analysis confirms that fraudulent postings disproportionately omit company logos and exhibit distinctive lexical patterns (e.g., "signing bonus") in their company profiles.