Knowledge Graph–Based Initialization for Alternating Least Squares: Bridging Semantic Gaps in Candidate–Job Matching

Abstract

Accurate candidate job matching on large recruitment platforms is challenged by severe data sparsity and the semantic heterogeneity of professional profiles. Although the Alternating Least Squares (ALS) algorithm is widely used due to its scalability, it relies solely on historical user–item interactions and does not account for semantic relationships between professional skills and occupations. This paper proposes a knowledge graph–based initialization strategy for the Alternating Least Squares (ALS) model that integrates collaborative behavioral signals with structured domain knowledge. Instead of conventional random initialization, the initial latent representations of candidates and job vacancies are constructed from skill embeddings derived from a professional knowledge graph. Experimental evaluation on a large-scale recruitment dataset demonstrates consistent improvements in recommendation quality. In cold-start scenarios, the proposed KG-ALS model achieves Recall@10 of 0.118 and NDCG@10 of 0.066, outperforming standard ALS with random initialization (0.071 / 0.038) and a neural Two-Tower retrieval model (0.092 / 0.053). This corresponds to an improvement of approximately 66% in Recall@10 and 74% in NDCG@10 compared to the baseline ALS model. Importantly, the proposed approach preserves the computational efficiency of classical matrix factorization, as the knowledge graph is used only during the initialization phase and does not affect inference complexity. The results indicate that incorporating semantic information into the initialization of latent factors can significantly enhance the effectiveness of matrix factorization models when operating on sparse interaction data.