Using artificial intelligence to identify administrative errors in unemployment insurance

Administrative errors in unemployment insurance (UI) decisions give rise to a public values conflict between efficiency and efficacy. We analyze whether artificial intelligence (AI) – in particular, methods in machine learning (ML) – can be used to detect administrative errors in UI claims decisions, both in terms of accuracy and normative tradeoffs. We use 16 years of US Department of Labor audit and policy data on UI claims to analyze the accuracy of 7 different random forest and deep learning models. We further test weighting schemas and synthetic data approaches to correcting imbalances in the training data. A random forest model using gradient descent boosting is more accurate, along several measures, and preferable in terms of public values, than every deep learning model tested. Adjusting model weights produces significant recall improvements for low-n outcomes, at the expense of precision. Synthetic data produces attenuated improvements and drawbacks relative to weights.