Probabilistic evaluation of enterprises' fraudulent report in pollutant levels based on machine learning: a case of hazardous waste generation quantity

Industrial enterprises are prominent sources of contaminants, and accurate reporting of contaminant discharge by these enterprises is essential for effective environmental management. However, fraudulent reporting, particularly underreporting hazardous waste (HW) generation, has become an intractable issue. HW is difficult to monitor automatically, and underreporting often implies improper treatment or illegal dumping, which can cause severe environmental risks. Regulators face challenges in detecting underreporting due to the lack of prior knowledge about the theoretical HW generation at the firm level. This study proposes a machine learning framework to detect fraudulent reporting by correlating HW generation with enterprises’ static characteristics and real-time manufacturing activities, such as wastewater discharge. The model was tested on two industrial sectors in Jiangsu, China, with datasets containing 10 % fraudulent samples. It achieved an accuracy of 42.9 %-55 % when observations with a probability of underreporting HW generation exceeding 90 % were flagged as fraudulent.