Predicting Salaries with Random-Forest Regression

Abstract

For companies it is essential to know the market price of the salaries of their current and prospective employees. Predicting such salaries is challenging, as many factors need to be considered, and large real datasets for learning are scarce. For this reason, research on salary predictions is comparably rare and limited. In this study, we investigate whether and how an advanced machine-learning approach, namely ensembles of random-forest regression, can achieve high-quality salary predictions. We use a large real dataset of more than three million employees and more than 300 professions. Our approach learns –for each profession– a random-forest regression model to predict salaries. In our evaluation, we show that this approach performs better than related work on salary prediction by machine-learning approaches with a mean absolute percentage error (MAPE) of 17.1%. We identify reducing the number of possible values of categorical variables, training separate models as well as outlier handling as the key factors for the results achieved.