Combining Large Language Model Embeddings and Graph Neural Networks for Accurate Job Recommendations

• Alexandre dos Santos Gualberto UFSCar
• Alan Demétrius Baria Valejo UFSCar
• Murilo Coelho Naldi UFSCar

Resumo

Finding the right job is hard for candidates and platforms alike. With countless postings, candidates struggle to discover relevant roles, while platforms fail to deliver precise matches. This inefficiency calls for smarter AI solutions beyond basic filters. This paper presents a hybrid framework for job recommendation that combines Large Language Model (LLM) embeddings with Graph Neural Networks (GNNs) to improve accuracy in matching candidates to job postings. The proposed model leverages GTE-1.5 to generate semantic embeddings from job descriptions and user profiles, augmented with geographic data, and integrates these with structural information from a bipartite user-job graph using GraphSAGE. The architecture includes a two-layer GraphSAGE network with BatchNorm and ReLU activation, followed by a dot-product decoder for link prediction. Evaluated on the Job Recommendation Challenge dataset, the hybrid model achieves an AUC of 0.97, outperforming text-only (AUC 0.75), structure-only (AUC 0.85), and classical matrix factorization (AUC 0.78) baselines. It also demonstrates superior performance in Precision@K and Recall@K across multiple thresholds. The results highlight the synergy between LLM-derived contextual understanding and GNN-based structural reasoning, offering a robust solution for large-scale job recommendation systems. Limitations and future directions, such as handling cold-start nodes and incorporating explicit location data, are also discussed.