Explainable Semantic Resume Screening With Fairness-Aware Candidate Ranking

Abstract

Large applicant pools make manual resume screening slow, inconsistent, and difficult to audit. Conventional Applicant Tracking Systems reduce effort through keyword filtering, but exact-term matching can miss semantically equivalent skills and offers limited justification for ranking decisions. This paper presents an explainable semantic screening framework that combines Sentence-BERT-style contextual similarity, structured qualification features, a validation-weighted ensemble of Logistic Regression, Random Forest, and Gradient Boosted Trees, and SHAP-based local explanations. Personally identifying information is separated from model features and retained only for authorized compliance workflows. Candidate ranking is driven by suitability probability, mandatory-skill coverage, and transparent feature contributions rather than raw keyword counts. A controlled synthetic benchmark of 1,760 resume–job pairs across eight role profiles is used to verify the analytical pipeline without presenting unverified real-world hiring outcomes. The proposed ensemble reaches 87.95% accuracy, 90.17% F1-score, and 94.90% ROC-AUC, improving over a keyword/skill baseline. Post-hoc group auditing shows a 1.51 percentage-point equal-opportunity gap in the controlled benchmark. The framework is intended as recruiter decision support, not an autonomous hiring authority, and emphasizes traceability, human override, and repeatable fairness monitoring.