This message has been cross posted to the following eGroups: Statistical Learning and Data Mining Section and Statistical Consulting Section .
-------------------------------------------
Sparse Logical Models for Interpretable Machine Learning
Friday, May 8, 2015
10:00 am to 12:00 pm Eastern
Information < http://www.enar.org/education/index.cfm>
Register < https://portal.enar.org/Events/SelectRegType.aspx?EventCode=WEB050815>
Presenter:
Cynthia Rudin, PhD, Associate Professor of Statistics, MIT CSAIL and Sloan School of Management, Massachusetts Institute of Technology
Description:
Possibly *the* most important obstacle in the deployment of predictive models is the fact that humans simply do not trust them. If it is known exactly which variables were important for the prediction and how they were combined, this information can be very powerful in helping to convince people to believe (or not believe) the prediction and make the right decision. In this talk I will discuss algorithms for making these non-black box predictions including:
- "Bayesian Rule Lists" - This algorithm builds a decision list using a probabilistic model over permutations of IF-THEN rules. It competes with the CART algorithm for building accurate-yet-interpretable logical models. It is not a greedy algorithm like CART.
- "Falling Rule Lists" - These are decision lists where the probabilities decrease monotonically along the list. These are really useful for medical applications because they stratify patients into risk categories from highest to lowest risk.
- "Bayesian Or's of And's" - These are disjunctions of conjunction models (disjunctive normal forms). These models are natural for modeling customer preferences in marketing.
- "The Bayesian Case Model" - This is a case-based reasoning clustering method. It provides a prototypical exemplar from each cluster along with the subspace that is important for the cluster.
I will show lots of applications of these models to healthcare and marketing.
Registration fees are by membership category, with a reduced fee for student members. Note that it is possible for non-ENAR members to participate, but you must create an account with the ENAR website to do so. As a benefit of ENAR membership, you may participate in one webinar per year at no charge.
The webinars are intended to be broadly available and ENAR encourages groups at your institution or workplace to participate together. ------------------------------
Lynn Eberly
University of Minnesota
------------------------------