Enhancing resource assignment efficiency in service industry: A predict-then-optimize approach with XGBoost

Abstract

This paper addresses the critical aspect of resource planning in a service context through an integrated predictive and prescriptive approach. Utilizing real-world data from a company providing repair and maintenance services, we demonstrate the use of an XGBoost model to forecast ad-hoc service demands and subsequently optimize resource assignment using a mathematical model. Our findings show that the prediction evaluation metrics significantly improve, highlighting the superiority of complex machine learning models over baseline models such as Linear Regression. Furthermore, the integration of the prediction into the decision-making process resulted in a 26.4% lower decision error compared to the baseline model. Our research also found that the deviations in prediction and optimal objective function values are not aligned. While the average error for MAE % in prediction is 22.2%, the error for the optimal objective function is much lower, reducing to 5.3%. However, although true for our case, this might not be generalizable. Furthermore, when comparing the baseline model with these results, it is also shown that an improvement in prediction accuracy also improves decision making error. Our results indicate that a combined predict-then-optimize approach outperforms the existing methods in both predictive and prescriptive performance, demonstrating its applicability in real-world scenarios.