Improving Inspection Resource Allocation to Control Organizational Misconduct

Machine learning is good at predicting where organizational misconduct is likely, but turning those predictions into real inspection decisions remains the hard part. We study how regulators with limited inspectors can schedule visits to prevent restaurant hygiene violations—a form of misconduct with serious public-health consequences.

Our prescriptive framework pairs a gradient-boosted tree model—trained on inspection history, time since the last visit, and Yelp review signals—with model-based planning that casts inspection scheduling as a capacity-constrained partially observable Markov decision process, solved with an easy-to-implement index policy. Tested on data from the Southern Nevada Health District (2012–2019), scheduling by predictions alone improves on fixed inspection intervals by 63%, and adding model-based planning delivers a further 48% gain. The approach extends naturally beyond restaurants to agencies such as OSHA, the FDA, and the EPA looking to do more with existing resources.