A new family of proximity graphs: class cover catch digraphs
Discrete Applied Mathematics
| Hi-index | 0.00 |
This proposal brings together two problems in classification. The first problem is how to design one of the simplest and oldest classifiers, the k-nearest neighbor classifier. The second problem is how to combine classifiers to produce a more effective classifier. Our immediate objective is to study a classifier that combines the predictions of a set of complementary nearest neighbor classifiers using several well-known machine learning algorithms. We use the term complementary to refer to a set of classifiers whose predictions may be combined to yield a classifier with accuracy higher than any of these component classifiers. The resulting architecture is an instantiation of the stacked generalization framework discussed by Wolpert [1992]. A central problem of this research is to characterize the senses in which classifiers are complementary and to present algorithms that create sets of complementary nearest neighbor classifiers. We propose algorithms that selectively incorporate different sets of prototypes into nearest neighbor classifiers. We bias our search for component nearest neighbor classifiers in favor of classifiers that incorporate only small sets of prototypes. In this proposal we provide evidence that a very small number of prototypes may be sufficient to give good generalization accuracy on several often used data sets. We also show that simple sampling and search algorithms with a stochastic component may be sufficient to find such prototypes. This proposal gives an introduction to the problems of nearest neighbor classifier construction and combination, reviews related research, gives an overview of the intended framework for stacked nearest neighbor classifiers, and introduces algorithms for nearest neighbor classifier construction and combination. Finally, we propose work to finish the dissertation.