Predição de links em grafos bipartidos para recomendação de empregos

Abstract

ABSTRACT This work develops and evaluates a recommendation system for job opportunities that integrates both textual and structural information, employing large-scale language models and graph neural networks. The main objective is to create a predictive framework that overcomes the limitations of traditional approaches by combining semantic representations obtained from textual embeddings with the ability of graph neural networks to capture interaction patterns within bipartite graphs. To achieve this, three distinct approaches were implemented: one based solely on text embeddings generated by language models, another founded exclusively on the graph structure, and a third that unifies both methodologies. The system was trained to predict links representing job applications using negative sampling, mini-batch training, and optimization strategies that ensured an efficient learning process. The results indicated that the hybrid approach achieved superior performance in terms of AUC, accuracy, precision, recall, and F1-Score, underscoring the importance of integrating semantic context with relational structure. Although the hybrid method presents challenges related to data quality and computational demands, these investments are justified by the improvements in recommendation accuracy. The study concludes that combining natural language processing techniques with graph-based learning constitutes a promising trategy for recommendation systems in complex environments, contributing to a better match between candidates and job vacancies and advancing recruitment processes