Skill-based clustering algorithm for online job advertisements

Clustering on the basis of categorical data is one of the challenging problems in data mining. The paper provides the clustering algorithm for job vacancies using information about the skills required. In the first step, the procedure of unstructured textual information standardization is proposed. The resulting procedures include stages of synonyms and general terms identification based on the combination of TF-IDF and $n$-grams approaches for translated and transliterated terms. Then, the algorithm is provided and validated on the data obtained from the cross-regional hiring platform. The algorithm provides validation of clusters’ extraction, including hierarchical cluster analysis and Girvan – Newman coalition search. Output number of clusters is verified with internal validity scores and suggests disjoint sets of terms that describe particular job occupation groups in the IT sector. Based on obtained clusters well-matched and mismatched terms are identified using Silhouette scores. Given procedures allow to minimize human involvement in clustering itself and produce reasonable clusters for the following interpretation and analysis. In general, the approach for clusters identification based on categorical data is provided and tested on a sample of online job advertisements. It has a high potential in use for feature engineering tasks in machine learning research and applied labor market research in economics.