Home – Home – Jornals – Scientific journal “Vestnik NSUEM” 2026 number 3

In a dynamically changing labor market with high volatility in job openings, standard job classifications that create intuitive job categories are often insufficiently detailed for in-depth economic and statistical analysis. A particular problem arises from labor market segments that are difficult to categorize using standard classifications, which, while accumulating a significant share of vacancies, defy simple classification, distorting the sectoral structure of labor demand. This article proposes and tests a method of contrasting multimodal graph clustering for detailing latent segments of regional labor markets. The methodology is based on a multi-stage approach involving normalization and vectorization of text data on job postings using the multilingual Sentence-BERT model, filtering semantic noise by calculating contrastive residual vectors relative to the centroids of known categories, enriching the feature space with multimodal data (salary level, work experience), constructing a weighted graph based on k-nearest neighbors and partitioning it into microclusters using the Leiden algorithm, followed by hierarchical aggregation of microclusters into macrosegments of a given size and balancing semantic proximity and weight constraints. The result of the work was the identification of interpretable latent segments within a heterogeneous category, often referred to as “other”: administrative and managerial personnel, blue-collar jobs, sales personnel, and highly specialized technical and production roles. The quality of the segmentation is confirmed by compliance with the specified weight constraints (target segment size of 5000 units) and the semantic interpretability of the resulting clusters based on c-TF-IDF. The proposed approach improves the analytical value of the data, automates the updating of the occupational classifier, and can be used to monitor structural shifts in the regional economy.