Analyzing the Impact of Class Imbalance Handling on Explainable Fake Job Posting Detection Using XGBoost and SHAP

Abstract

Fake job postings on online recruitment platforms can cause job seekers to suffer financial losses and identity theft. The detection task for such fraudulent postings has a core challenge: datasets suffer from severe class imbalance, where fake postings account for only a tiny fraction of the total data. Most previous studies only focus on models’ classification performance, and rarely discuss the impact of class imbalance processing on feature attribution and model interpretability. This study adopts the XGBoost and SHAP methods to conduct detection research. The framework built for this study first completes text preprocessing, then extracts hybrid features by combining TF-IDF and metadata attributes, and evaluates four class imbalance processing strategies in total: Baseline, SMOTE, Borderline-SMOTE, and ADASYN. Experimental results show that compared with the baseline model, oversampling methods improve the detection performance for the minority class. ADASYN delivers the best performance, with corresponding scores of 79.23% for Recall, 81.91% for F1-score, and 88.70% for G-Mean. SHAP analysis finds that the model’s feature attribution pattern changes, with its attention shifting to fraud-related features, while hascompanylogo consistently remains the feature with the highest influence. This study confirms that class imbalance processing affects both classification performance and model interpretability