This study compares the effects of using various model classes and different feature sets for the prediction of labor market chances of young unemployed. For this comparison the original Austrian labor market algorithm AMAS, which is based on a logistic regression und trained upon administrative data, is evaluated on a dataset of young unemployed in Vienna and then compared to additional model classes and various survey-based covariate sets as a benchmark in regard to different performance and fairness measures. Overall, choosing a diverse set of variables using not only administrative data but also survey data with multiple variable categories, like behavior, work attitudes and personality, did improve performance accuracy and fairness over all model classes. Nevertheless, a general trade-off between overall accuracy and fairness was also observed, consistent with numerous other publications. On average, the model class "extreme gradient boosting" performed best in regard to performance and fairness over all tasks.
Whole text: Mather's Thesis
