Predictive hiring leverages machine learning to forecast candidate success, yet existing approaches suffer from two limitations: reliance on single-method feature selection that lacks robustness, and sensitivity to neural network initialization that impairs convergence. This study introduces two contributions integrated into a unified framework. First, Stability-Weighted Multi-Criteria Feature Selection (SW-MCFS) is proposed, which aggregates four heterogeneous scoring methods—Mutual Information, Wald statistical significance, Fisher discriminant loading, and Permutation Importance—through a cross-validation stability-weighted consensus function. Unlike single-method approaches, SW-MCFS weights each method proportionally to its ranking consistency across folds, producing robust and data-driven feature subsets. Second, Adaptive Particle Swarm Optimization (APSO) is introduced, a PSO variant featuring fitness-landscape-aware inertia adaptation and Levy flight perturbation for stagnation escape. The framework is evaluated on 10, 247 recruitment records from a North African telecommunications company and benchmarked against Random Forest, XGBoost, SVM, standard ANNs, and classical LR/DA-based approaches through 10-fold cross-validation. The integrated SW-MCFS-APSO-SGD framework achieves 76.8% accuracy, significantly outperforming XGBoost (73.8%, p = 0.012), standard PSO-SGD (75.2%, p = 0.041), and LR-based feature selection (74.6%, p = 0.028). Ablation studies confirm that SW-MCFS contributes 1.6%accuracy gain over single-method selection, while APSO improves performance by 0.8% with 31% faster convergence compared to standard PSO. SHAP analysis reveals communication skills, experience, and seniority as dominant predictors with minimal demographic influence. It is noted that the accuracy ceiling may partly reflect inherent label noise in subjective performance assessments. The proposed framework demonstrates effectiveness on organizational recruitment data, warranting further cross-domain validation to establish broader generalizability.
Khallouk et al. (Thu,) studied this question.