This study aims to improve the quality metrics of predicting the severity of psychological characteristics of online social network users using machine learning models by forming new features through their transformation from existing ones using multidimensional analysis methods and manual data transformation. The initial features are extracted data from users’ pages on online social networks, and the results of users’ Big Five personality tests are used as the training target. Two machine learning models were used in the study: CatBoost and RandomForestRegressor, which were compared after training on various features: initial and generated. The results showed that synthetic feature generation based on existing features improves the result by an average of 34.17% in terms of MAE and 22.24% in terms of RMSE.