OCA-XHireNet: A leakage-controlled evaluation of graph learning, explainability and fairness optimization for recruitment

Automated recruitment systems often rely on keyword screening and opaque scoring, raising concerns about semantic adequacy, transparency, and fairness. This paper presents OCA-XHireNet (Organizational-Context-Aware eXplainable Hiring Network), an eight-module framework that couples resume representation learning, knowledge-graph grounding, graph neural refinement, post-hoc explanation, and multi-objective fairness optimization, and evaluates its components on public data with protocols that control for duplicate leakage, fold dependence, class weighting, and overfitting. An audit of the Kaggle Resume Dataset showed that only 166 of its 962 records are unique; record-level splitting inflated matching accuracy above 99%, whereas on the unique resumes TF-IDF with a random forest (75.3%) and averaged pretrained word vectors (74.5%) were statistically indistinguishable and significantly outperformed a frozen RoBERTa-base encoder (37.7%) and a from-scratch transformer (30.6%). Graph refinement with GCN or GAT did not improve on the encoders’ own classifiers in any tested configuration and significantly reduced accuracy with both pretrained representations. On the IBM HR Analytics attrition data, Kamiran–Calders reweighing lowered the demographic parity difference by 19% (not significant) while nearly halving the selection rate, at a significant F1 cost; with class weighting matched, it did not reduce disparity. Across ten seeds, NSGA-II fronts attained a higher held-out hypervolume than the single-objective baseline (0.107 versus 0.081), but near-zero disparity on the optimization data did not transfer reliably. TreeSHAP ranked overtime, monthly income, and stock options as the largest contributors to predicted attrition, and ESCO v1.0 skill labels occurred in 91.6% of unique resumes but did not improve classification.