Data-Driven Capacity Management with Machine Learning: A Novel Approach and a Case-Study for a Public Service Office

Abstract

In this paper we consider the case of a public service office in Germany that provides services such as handling passports and ID card applications, notifications of change of addresses, etc. Their decision problem is to determine the staffing level for a specific staffing time-slot (e.g., next Monday, 8 am–12.30 pm). Required capacity is driven by features such as the day of the week, whether the day is in school vacations, etc. We present an innovative data-driven approach to prescribe capacities that does not require any assumptions about the underlying arrival process. We show how to integrate specific service goals (e.g., “At most 20% of the customers should have to wait more than 20 min”) into a machine learning (ML) algorithm to learn a functional relationship between features and prescribed capacity from historical data. We analyze the performance of our integrated approach on a real-world dataset and compare it to a sequential approach that first uses out-of-the-box ML to predict arrival rates and subsequently determines the according capacity using queuing models. We find that both data-driven approaches can significantly improve the performance compared to a naive benchmark and discuss benefits and drawbacks of our approach.