This dissertation deals with applying machine learning and mathematical optimization methods for data-driven decision-making. It contributes to research on building machine learning models that capture human behavior and human preferences and that can be used to improve operations and optimization processes. In the field of churn prevention, churn prediction models have primarily been trained on one time slice of data. This work evaluates an approach to train models on data from multiple time slices and identifies two effects that contribute to an improvement in predictive performance: an increase in sample size as well as training on samples from different time slices. The multi-slicing approach makes models more generalizable under changing conditions and allows for predictions that are more accurate. In a field experiment with B2B customers of a convenience wholesaler, this thesis demonstrates how customer churn can be decreased by basing targeting decisions and retention efforts on predicted churn probabilities. In the field of crew scheduling, this work shows how benefits from machine learning and optimization can be combined to deliver better solutions for complex planning problems. For a railway freight carrier, feedback from planners regarding crew schedules is used to train a machine learning model. This thesis introduces an approach to integrate predicted planner feedback into the optimization process for improving the expected planner acceptance of solutions.weiterlesen