Machine Learning/Public Policy Proposal
- Hamburg Hall
- WILLIAMS HERLANDS
- Ph.D. Student
- Joint Ph.D. Program in Machine Learning and Public Policy
- Carnegie Mellon University
Change modeling for understanding our world and the counterfactual one(s)
Detecting, analyzing, and modeling changes provide essential information for understanding scientific processes and human behavior. While change analysis is fundamental in machine learning and statistics, many standard models are limited in expressiveness or make unrealistic simplifying assumptions. This thesis focuses on two interrelated elements of change analysis.
First, we provide rich characterization of changes by developing new methods for modeling complex changes and for detecting anomalous patterns in real world data. In order to characterize a change we automatically model the ``null'' regions of stability in the data and identify where "alternative" regions of change or anomalies exist. By modeling how the alternative regions differ or evolve from the null regions, we show that we are able to use that information for scientific discovery and for early event detection.
Second, we consider causal and counterfactual inference by exploiting changes to uncover the generative structure of data. By isolating changes in data we can reason about what would have occurred in the absence of a change. Such reasoning enables us to predict the counterfactual world and estimate the causal impact of certain variables or interventions in a data set.
Using ten different public interest data sets we employ our methods to characterize changes and identify causal mechanisms that can provide scientific and policy relevant insights. Specifically, we concentrate on health policy and urban data, much of which exhibit distinct spatial and demographic patterns. The data we explore includes measles incidence, health insurance usage rates, water lead testing, requests for municipal services, urban opioid deaths, weather related damage in urban neighborhoods, and urban school absenteeism.
Daniel Neill (Chair)
Andrew Gordon Wilson (Cornell University)