Interpretable Statistical Learning

Restricted (Penn State Only)
Author:
Jones, Zachary Maddox
Graduate Program:
Political Science
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
June 08, 2017
Committee Members:
  • Bruce A Desmarais Jr., Dissertation Advisor
  • Bruce A Desmarais Jr., Committee Chair
  • Bumba Mukherjee, Committee Member
  • Christopher Jon Zorn, Committee Member
  • Bharath Kumar Sriperumbudur, Outside Member
Keywords:
  • machine learning
  • interpretability
Abstract:
Statistical learning methods, which are a flexible class of methods capable of estimating complex data generating functions, offer an attractive alternative to conventional methods, which often rely on strong, frequently inappropriate, assumptions about the functional form of the data generating process. A key impediment to the use of statistical learning methods is that they often output a "black box" which makes predictions but cannot be directly interpreted. I show that this need not be the case by describing and implementing methods based on Monte-Carlo integration that are capable of making any method that generates predictions interpretable. This allows researchers to learn about relationships in the data without having to prespecify their functional form. I illustrate this approach using a simulated example, a proposed application to the prediction of burglary reports in Chicago, and an application to the study of political violence.