Data Quality-Aware Detection of Fraudulent Job Advertisements Using Text and Missing-Metadata Patterns

Abstract

Fraudulent job advertisements can expose applicants to financial loss, identity theft, and wasted effort. Most automated detectors emphasize wording, although an advertisement’s missing or incomplete metadata may provide complementary evidence about how it was created. This paper proposes a data quality-aware decision-support framework that joins term frequency–inverse document frequency (TF–IDF) text features with explicit missingness indicators, aggregate completeness measures, and simple text-quality variables. A reproducible CPU-only simulation uses the open Employment Scam Aegean Dataset and a fixed stratified holdout. After removing 330 repeated predictor-identical records, the text-only logistic model obtained precision 0.673, recall 0.830, F1-score 0.743, balanced accuracy 0.905, and precision–recall area 0.877. Fusion increased recall to 0.889, balanced accuracy to 0.925, and precision–recall area to 0.886, but reduced precision to 0.535 and F1-score to 0.668. Metadata quality alone was considerably weaker. Thus, missingness provided a modest recall-oriented signal rather than a universal improvement. The practical contribution is a transparent ablation protocol for deciding when completeness patterns warrant inclusion in human-reviewed recruitment-fraud triage.