Date of Award

2-4-2025

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Graph Neural Networks; Job; Job-Resume Compatibility; Large Language Models; Resume

Supervisor

Dan Wu

Rights

info:eu-repo/semantics/openAccess

Abstract

In today’s fast-moving hiring landscape, traditional Applicant Tracking Systems (ATS) rely heavily on keyword matching to screen resumes. This approach can miss qualified candidates and lead to poor matches, showing the need for smarter and more effective screening methods. This study introduces a compatibility scoring model using a graph neural network (GNN) and a large language model (LLM) to score how well the resume matches the job description. The model is built on a real-world dataset where each job listing is paired with multiple resumes, and HR experts score these resume-job pairs, providing valuable training data. To create the input for the GNN, the text from resumes and job descriptions is converted into a graph, where nodes represent key sections such as skills, experience, and education, and edges capture relationships like proficiency levels for skills. Each node and edge is further furnished with embeddings generated by the Large Language Model (LLM), which enriches the graph with detailed semantic and contextual information. These embeddings enable the GNN to process the graph effectively by learning complex patterns and relationships between nodes and edges. The use of LLM not only transforms textual information into the numerical format required for graph construction but also captures deeper semantic meaning and contextual nuances, further enhancing the GNN’s ability to make accurate predictions. The proposed model not only scores the compatibility between resumes and job descriptions but also explains how the scoring is determined, making the process transparent. Job seekers receive clear feedback on the alignment of their resume with a job description, with actionable improvement tips. Recruiters gain more effective tools to evaluate candidates. This novel method, which combines the graph neural network (GNN) and the large language model (LLM), improves the assessment of the compatibility score by 25\% compared to traditional keyword-based methods, aligning more closely with HR expert evaluations. This highlights the model's ability to handle real-world hiring complexities, bringing us closer to fairer, smarter, and more efficient hiring practices.

Download

Share

COinS