Interplay Between Statistical Learning and Decision Making

Tuesday, March 4, 2014 - 4:15pm to 5:15pm
3:45pm in 32-G882
Cynthia Rudin, MIT

We often use predictive models to make decisions afterwards. For instance, we might estimate the number of patients at a medical clinic and then schedule staff to serve those patients. Each decision comes at a cost (for instance, the cost of staffing the clinic). At the same time, the class of "good" predictive models might be quite large (called the "Rashomon effect" by Breiman) leading to uncertainty in our decisions. This bring into question how we should make decisions, and what are our operational costs.

In this talk I will discuss joint work with EECS PhD student Theja Tulabandhula linking statistical learning theory to decision making. We aim to answer 3 questions:

1) For the range of "good" predictive models, what is the range of operational costs? (We want to know how much it may cost to solve our problem.)

2) If we have prior knowledge about the operational cost, can we use this to help in prediction? This question is not just theoretical - usually our prior knowledge is about norms of coefficients of linear models, but managers do not generally think in terms of norms. They may think in terms of operational costs instead.

3) If we were to use robust optimization to create our solution, can we learn the uncertainty set from data? (We want our decision to be feasible with high probability.)

This talk is on the conceptual side, unconventional, and definitely off-beat.