A Heterogeneous Bipartite Graph Neural Network for Bidirectional Recommendation in Online Recruitment

Recruitment on online platforms is a data‐driven bilateral decision‐making process in which job seekers and recruiters interact within a large and complex socio‐technical system. A useful recommendation should therefore serve both sides simultaneously, yet most existing studies address only one direction or rely on a single representation source, modelling neither the heterogeneous preferences of the two parties nor the interaction structure that connects them. To address this limitation, we present HBGBR, a heterogeneous bipartite graph neural network for bidirectional recruitment ranking. CNN‐based text encoders extract semantic representations from large‐scale resume and job texts, and a heterogeneous bipartite graph models resumes, jobs, and their observed interactions. A layerwise adversarial module then aligns the cross‐domain graph representations before text and graph features are fused into a unified compatibility score. Bidirectional relevance is evaluated by ranking jobs for candidates and candidates for jobs. Experiments on a real‐world dataset show that HBGBR consistently outperforms 12 representative baselines across five categories on NDCG@5, MRR@5, MAP@5 and HIT@1 in both directions, with improvements of up to 15.29% in candidate‐side HIT@1 over the strongest baseline. User‐level paired Wilcoxon tests confirm that all improvements remain statistically significant after Bonferroni correction. Ablation studies further validate the necessity of adversarial alignment and layerwise propagation for achieving robust bidirectional performance.