This paper addresses the lack of tools for automated assessment of individuals’ professional interests. The study aims to improve the reliability of Holland code (RIASEC) reconstruction using digital traces from online social networks through a two-stage machine learning framework. To this end, a method is proposed that simultaneously solves two complementary tasks: (1) predicting the dominant professional type (Top-1), and (2) reconstructing the top three professional types (Top-3). The feature set includes quantitative indicators of user activity on the social network, categorical variables (such as gender and status), and a normalised frequency vector of subscription topics. Ensemble algorithms (Random Forest, ExtraTrees, LightGBM, CatBoost) and a neural network model trained with the ListNet loss function were employed for model development. The final solution adopts a two-headed ensemble architecture, in which the ExtraTreesClassifier predicts the Top-1 type and the RandomForestRegressor reconstructs the Top-3 vector. The proposed method demonstrates high predictive performance even under conditions of rank ambiguity and ties, underlining the novelty of the approach. The resulting system can be integrated into personalised recommendation tools to support faster, data-driven career guidance.