Using Artificial Intelligence Techniques for Evidence-Based Decision Making in Government: Random Forest and Deep Neural Network Classification for Predicting Harmful Algal Blooms in New York State

Abstract

Advances in artificial intelligence techniques have shown tremendous potential as a decision support tool for government agencies. However, recent studies typically highlight the combination of large-scale datasets and high-performance computing technologies, which is frequently far away from the reality of many public agencies that still heavily rely on their legacy systems and labor-intensive practices. Using that non-ideal organizational and technical infrastructure, they have to face very complex problems and propose policy solutions. Harmful algal blooms (HABs) are one of such problems. HABs have increasingly become a serious environmental issue in the United States. However, the rapid and sporadic growth of algae and the current standard relying on manual sampling weaken the agencies' response. To overcome this limitation, in this study, we attempt to bridge advanced AI technologies and current government practice by examining the potential of artificial intelligence by comparing the performance of linear probability, random forest, and deep neural network algorithms in predicting HABs with manual sampling data. By integrating manually-sampled HABs data with predictors from publicly-available datasets (land use, weather, and drought), we demonstrate that random forest and deep neural network (DNN) algorithms improve the specificity of the prediction, increasing the true negative rate. Albeit not ideal, we believe that this approach can benefit public agencies that are forced to respond to HABs with limited resources for investing in improving legacy systems. Accurate prediction with limited data could still be useful for certain government decisions, even when the mechanisms of causality are not totally clear. AI techniques have the potential to improve these predictive capabilities.