In recent years, researchers have made considerable progress on the worst-case analysis of inductive learning tasks, but for theoretical results to have impact on practice, they must deal with the average case. In this paper we present an average-case analysis of a simple algorithm that induces one-level decision trees for concepts defined by a single relevant attribute. Given knowledge about the number of training instances, the number of irrelevant attributes, the amount of class and attribute noise, and the class and attribute distributions, we derive the expected classification accuracy over the entire instance space. We then examine the predictions of this analysis for different settings of these domain parameters, comparing them to experimental results to check our reasoning.
[1]
M. Levine.
Hypothesis behavior by humans during discrimination learning.
,
1966,
Journal of experimental psychology.
[2]
M. Kearns,et al.
Recent Results on Boolean Concept Learning
,
1987
.
[3]
Michael J. Pazzani,et al.
Average Case Analysis of Conjunctive Learning Algorithms
,
1990,
ML.
[4]
Pat Langley,et al.
Machine learning as an experimental science
,
2004,
Machine Learning.
[5]
N. Littlestone.
Learning Quickly When Irrelevant Attributes Abound: A New Linear-Threshold Algorithm
,
1987,
28th Annual Symposium on Foundations of Computer Science (sfcs 1987).
[6]
Daniel S. Hirschberg,et al.
Average case analysis of a k-CNF learning algorithm
,
1991
.