Comparative Study of Machine Learning Techniques in Prediction of Superfund Sites

Superfund sites are contaminated regions caused by the mismanagement of hazardous waste dumping. Their existence poses important questions regarding their locations, given that research has shown that poorer communities tend to have disproportionately higher numbers of hazardous waste sites. This project aims to investigate the validity of this correlation by comparing predictive performance across various approaches with data from the 2010 U.S. Census and Superfund site locations from the National Priorities List. To account for a large, imbalanced dataset, the data was subjected to preprocessing methods like resampling and feature selection to optimize for model accuracy. Two separate model training pipelines were implemented: one in Python-based notebooks and the other in the IBM Watson platform. Afterward, a comparative study of the results of the two pipelines and their individual models was performed. In the findings, both Jupyter notebook-trained models and their IBM counterparts achieved accuracies greater than 97% on Superfund site prediction and F1 values that reflect optimal overall classification performance. The study also found that features highly correlated to the presence of Superfund sites include the number of people living in subsidized housing units, the number of people eligible to participate in the labor force, and addresses with bilingual mail. This research aims to raise awareness about environmental discrimination in the United States and can be extended to incorporate data from sites of a similar nature in other countries through respective data releases.