Increasing Inspection Efficiency using Machine Learning
Facility inspections represent a key enforcement practice across regulatory domains, but are costly and infrequent. Despite increasingly scarce government resources, regulators can improve compliance by adopting data-driven inspection targeting practices and identifying violators more efficiently. Technological advances in computing capabilities, big data availability, and analytic methods can transform the way agencies implement environmental enforcement by providing reliable predictions of which facilities are most likely to be in violation.
Results: 82% increase in detection of hazardous waste violations, with no cost increase
Since 2017, E&E Lab researchers have been working onsite at U.S. EPA headquarters in Washington, D.C. to test the power of machine learning to improve enforcement targeting. The E&E Lab has developed a robust model to harness administrative data spanning nearly two decades to predict which facilities are most likely to violate hazardous waste regulations. Although the model substantially exceeded the EPA’s historical hit rate, regulators remained unconvinced, so we proposed a figurative horse race – pitting model-chosen inspections by EPA program staff. Results to date suggest the model yields a 82% increase in detection of serious hazardous waste violators compared to the EPA’s status quo practices, with no increase in cost.
This model is the first EPA tool driven by machine learning and the first EPA enforcement practice to be validated through a modern evaluation. Driven by this success, the EPA has decided to scale this approach nationwide and to launch a National Targeting Center, which will offer predictive analytics models to improve targeting as a cornerstone service to environmental regulators at the regional and state level. We are now developing new models to target inspections of facilities regulated under the Clean Air Act, and the Clean Water Act, which will leverage advanced remote sensing and modeling technology with potential to drastically improve targeting efficiency