|
Activity Number:
|
241
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, July 31, 2007 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
WNAR
|
| Abstract - #309234 |
|
Title:
|
Locally Optimal Weighting and Classifier Selection in Ensembles
|
|
Author(s):
|
Melissa Fazzari*+ and Hongshik Ahn and Wei Zhu
|
|
Companies:
|
Stony Brook University and Stony Brook University and State University of New York at Stony Brook
|
|
Address:
|
24 Forest Drive, Centerport, NY, 11721,
|
|
Keywords:
|
ensembles ; classification ; bias-variance decomposition ; nearest neighbors
|
|
Abstract:
|
Ensemble-based classification improves generalization accuracy through a mix of variance and bias reduction. The largest gains are found through the aggregation of strong, but diverse classifiers. For each individual test point, the best set of classifiers and their ensemble weights may be highly varied. We examine locally optimal classification, a weighting scheme for combining classifiers based on local performance. Weights are determined based on a bootstrap estimate of variance for each classifier at each training point. Classification of test instances is achieved by combining the predictions across classifiers using the weights of the training set nearest-neighbors. Other methods of ensemble selection and combination are also explored, including low bias combining and a correlation-based approach. Variable importance across all ensemble members is examined and summarized.
|
