Resume-Job Compatibility Scoring Using Graph Neural Networks and Large Language Models

Abstract

Traditional Applicant Tracking Systems (ATS) rely heavily on keyword matching, often overlooking semantically relevant resumes due to phrasing differences or formatting issues. To address this limitation, we propose a resume-job compatibility scoring framework that integrates Graph Neural Networks (GNNs) with Large Language Model (LLM)-generated embeddings. Each resume-job pair is represented as a graph where nodes encode skills, education, and experience, and edges capture relationships such as proficiency levels. To enrich the semantic representation of each entity, contextual embeddings are generated using pre-trained LLMs. These enriched graphs are processed by a GNN to predict compatibility scores. Experiments on a real-world HR-annotated dataset demonstrate that our model achieves a 25% improvement in predictive accuracy over traditional keyword-based baselines. In addition to accuracy, the model offers interpretability through attention-based mechanisms, providing actionable feedback for candidates and insights for recruiters. Our findings suggest that combining LLMs with GNNs can significantly enhance fairness, transparency, and efficiency in automated hiring systems.