Knowledge graph embedding with semantic similarity for e-recruitment recommendation systems

This paper proposes an e-recruitment recommendation framework to improve matching between job offers and candidate resumes. It relies on a knowledge graph built from an ontology structured around Domain, Occupation, and Skill, aligned with ESCO and populated using 300,000 profiles. The framework focuses on knowledge graph embeddings (KGE) combined with semantic similarity modeling to improve the matching between candidates and job offers. It uses a semantic similarity model based on a transformer autoencoder (TSDAE) to capture contextual relationships between entities, including skills and occupations. These semantic representations are integrated with KGE-based candidate ranking, enabling the system to consider both explicit and inferred skills from professional experience, and thus provide more accurate and context-aware recommendations. Evaluation results show promising performance: expert validation of the knowledge graph achieved a 70-74% skill validation rate in IT domains. The tsdae-distiluse-multilingual-bert model performed best in similarity tasks, while the ConvE based ranking approach slightly outperformed DistMult with mp@3 = 0.775 and ndcg@3 = 0.810.