Graph Neural Networks for Candidate-Job Matching: An Inductive Learning Approach

Abstract This work introduces a novel graph-based approach to candidate-job matching using Graph Neural Networks (GNNs). We analyzed data from 62 real-world selection processes encompassing 8360 unique candidates and 9532 applications, characterized by extreme class imbalance (95% rejection rate). Our methodology constructs purpose-built bipartite graphs for each candidate-job pair, with 14 nodes representing candidates, jobs, and their respective attributes extracted using Large Language Models. Each graph contains a minimum of 15 edges representing semantic relationships between entities, with edge weights derived from embedding similarity measures. We empirically evaluated five GNN architectures (GCN, MIGNN, GIN, GAT, GraphConv) against standard neural networks across binary and ordinal classification tasks. In binary classification, graph-based approaches consistently outperformed non-graph baselines, with GCN achieving 65.4% balanced accuracy compared to 55.0% for the MLP baseline. GNN models also demonstrated superior minority class detection, with GCN correctly identifying 48.9% of qualified candidates versus only 8.5% for MLP. Statistical analysis revealed that higher recruitment stages correlate with increased graph connectivity, validating our graph construction methodology. While all models struggled with ordinal classification, the explicit modeling of semantic relationships through graph structures enabled effective binary discrimination for candidate screening, offering a promising direction for augmenting human decision-making in recruitment processes while maintaining interpretability.