Порівняльний аналіз моделей машинного навчання в задачі передбачення вигорання співробітників

Abstract

The article explores the problem of predicting the emotional burnout syndrome of employees , which is relevant due to the high level of stress in the modern world. The study uses the publicly available dataset "Are your employees burning out" from the competition on the HackerEarth platform. A comparative analysis of three traditional machine learning models based on classical machine learning approaches (linear regression, Random Forest, XGBoost) and three Bayesian models (Bayesian linear regression, varying intercept model, varying intercept and slope model) was carried out in the study. The change in the quality of the models is studied for different sizes of data sets, ranging from 13,000 (i.e., the full training set, which accounted for 70% of all data) to 25 observations, including testing on the full data set. It is demonstrated that XGBoost is the best model for large data sets. However, when the training sample size is reduced to less than 5000 observations, the validation performance of the XGBoost model becomes significantly less accurate and becomes lower than the corresponding metrics for Bayesian models. After optimizing such hyperparameters as tree depth, number of trees, learning rate, and others, the quality of XGBoost improved significantly, but did not make it stable enough to demonstrate better results than Bayesian models on samples of less than 600 observations. Bayesian models, on the other hand, in addition to being better on small samples, also allow estimating the "confidence" in the predicted values, which is an important feature for a specific tasks. However, they also have a significant disadvantage in the form of much greater computational complexity, which leads to an increase in training time. In conclusion, results of this study emphasize the importance of careful selection of a model that considers the peculiarities of the amount and quality of available data. Bayesian models have proven to be highly effective with a small amount of data, due to their ability to consider uncertainty and insufficient information.