Exploiting the Absence of Irrelevant Information: What You Don't Know Can Help You
Most inductive inference algorithms are designed to work most effectively when their training data contain completely specified labeled samples. In many environments, however, the person collect- ing the data may record the values of only some of the attributes, and so provide the learner with only partially specified samples. This can be mod- eled as a blocking process that hides the values of certain attributes from the learner. While block- ers that remove the values of critical attributes can handicap a learner, this paper instead focuses on blockers that remove only super attribute values, i.e., values that are not needed to clas- sify an instance, given the values of the other unblocked attributes. We first motivate and for- malize this model of super blocking?, and then demonstrate that these omissions can be quite useful, showing that a certain class that is seemingly hard to learn in the general PAC model | viz., decision trees | is trivial to learn in this setting. We also show how this model can be ap- plied to the theory revision problem.
Citation
R. Bharat Rao,
R. Greiner,
T. Hancock.
"Exploiting the Absence of Irrelevant Information: What You Don't Know Can Help You". November 1994.
| Keywords: |
Relevance, irrelevance, missing information, PAC, machine learning |
| Category: |
In Workshop |
BibTeX
@misc{BharatRao+al:94,
author = {R. Bharat Rao and Russ Greiner and Thomas Hancock},
title = {Exploiting the Absence of Irrelevant Information: What You Don't
Know Can Help You},
year = 1994,
}Last Updated: July 07, 2007
Submitted by Russ Greiner 
