Learning Interpretable Decision Functions via Multi-Task Contextual Policy Recovery

In this paper, we propose a new approach to learning decision policies that balances accuracy with interpretability. Existing models often sacrifice one for the other, leading to difficulties in understanding and improving medical decision-making processes. Our proposed method, Contextual Policy Recovery (CPR), reframes the problem of policy learning as a multi-task learning exercise, where each task is associated with a specific context. This allows us to learn context-specific policies that can be easily understood and interpreted.
Imagine you’re trying to build a recipe for your favorite dish. You want to make sure it turns out great, but you also want to know exactly what ingredients are going into each bite. Existing decision-making models are like a generic recipe book that doesn’t tell you much about the individual ingredients or how they interact. CPR is like having a personal chef who can give you a detailed list of ingredients and explain why they’re important for your dish to turn out right.
Our approach uses black-box models, which are like having a secret ingredient that only the chef knows about. These models are trained on large datasets and learn how to make decisions based on patterns and trends. But sometimes, you want to know more than just the pattern – you want to understand why the chef made that decision in the first place. That’s where glass-box models come in, which are like having a cookbook that tells you exactly how the chef prepared each dish. Glass-box models provide a detailed explanation of how decisions were made, and they can help identify biases or errors in the decision-making process.
By combining black-box and glass-box models, CPR provides both accuracy and interpretability in decision-making. It’s like having a recipe book that tells you exactly what ingredients to use and why, while also giving you a detailed explanation of how each dish was prepared. This approach has already gained attention in the medical community for improving standards of care by detecting biases and explaining suboptimal outcomes.
In summary, CPR is a new approach to learning decision policies that balances accuracy with interpretability. By framing policy learning as a multi-task learning exercise, we can learn context-specific policies that are both accurate and easy to understand. This allows medical professionals to improve decision-making processes while also gaining a better understanding of how those decisions were made.