Strategy for preventing human trafficking through verification of online job vacancies in Indonesia: English

Abstract

This study addresses the rise of online job ads used to recruit victims of human trafficking (TPPO). We propose a practical screening approach that combines automated checks with human moderation. The goal is not to prove crimes, but to prioritize high-risk ads for fast review and referral. Using a public dataset of 500 job postings (fake_job_postings_500), we clean the text and basic metadata, extract simple text features (TF–IDF), and add light verification signals (e.g., contact and firm consistency). We then train two models in a leakage-safe pipeline: calibrated Logistic Regression (LR-cal) and Random Forest (RF). Performance is evaluated with standard accuracy measures ROC-AUC, PR-AUC, F1 plus calibration (how well risk scores match reality) and triage metrics that reflect real operations: precision for the highest-risk group, recall for all medium-and-above risk, and the share of ads moderators must review. Results show LR-cal is accurate and well-calibrated (5-fold means: ROC-AUC 0.993, PR-AUC 0.986, F1 0.934). In triage with thresholds T_high = 0.80 and T_med = 0.50, LR-cal yields Precision@High = 1.00 and Recall@≥Med=0.925 with ~34% of ads needing review. RF reaches near-ceiling accuracy (1.00/1.00 at ~35.3% workload) but requires careful calibration and leakage auditing. Practical contribution: AI-assisted, risk-based gatekeeping can reduce exposure to Human Trafficking or TPPO at the source. We recommend: (1) adopting calibrated models with adjustable thresholds; (2) standard operating procedures (SOPs) for cross-platform verification, including Know Your Customer (KYC) and Open-Source Intelligence (OSINT) checks; and (3) direct integration with official reporting channels to escalate flagged ads swiftly.