Addressing two issues in machine learning : interpretability and dataset shift
Name
1126649834-MIT.pdf
Size
18.2 MB
Format
Adobe PDF
Checksum (MD5)
6525025816819e2d21b865b0791d3715
Author(s)
Wang, Fulton.
Advisor(s)
Cynthia Rudin.
Date Issued
2018
Publisher
Massachusetts Institute of Technology
Abstract
In this thesis, I create solutions to two problems. In the first, I address the problem that many machine learning models are not interpretable, by creating a new form of classifier, called the Falling Rule List. This is a decision list classifier where the predicted probabilities are decreasing down the list. Experiments show that the gain in interpretability need not be accompanied by a large sacrifice in accuracy on real world datasets. I then briefly discuss possible extensions that allow one to directly optimize rank statistics over rule lists, and handle ordinal data. In the second, I address a shortcoming of a popular approach to handling covariate shift, in which the training distribution and that for which predictions need to be made have different covariate distributions. In particular, the existing importance weighting approach to handling covariate shift suffers from high variance if the two covariate distributions are very different. I develop a dimension reduction procedure that reduces this variance, at the expense of increased bias. Experiments show that this tradeoff can be worthwhile in some situations.
Description
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2018
Cataloged from PDF version of thesis.
Includes bibliographical references (pages 71-77).
Subjects
Electrical Engineering and Computer Science.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
MIT theses are protected by copyright. They may be viewed, downloaded, or printed from this source but further reproduction or distribution in any format is prohibited without written permission.
Persistent DSpace Link
