A comparative study of data sampling techniques for constructing neural network ensembles.

Ensembles with several classifiers (such as neural networks or decision trees) are widely used to improve the generalization performance over a single classifier. Proper diversity among component classifiers is considered an important parameter for ensemble construction so that failure of one may be compensated by others. Among various approaches, data sampling, i.e., different data sets for different classifiers, is found more effective than other approaches. A number of ensemble methods have been proposed under the umbrella of data sampling in which some are constrained to neural networks or decision trees and others are commonly applicable to both types of classifiers. We studied prominent data sampling techniques for neural network ensembles, and then experimentally evaluated their effectiveness on a common test ground. Based on overlap and uncover, the relation between generalization and diversity is presented. Eight ensemble methods were tested on 30 benchmark classification problems. We found that bagging and boosting, the pioneer ensemble methods, are still better than most of the other proposed methods. However, negative correlation learning that implicitly encourages different networks to different training spaces is shown as better or at least comparable to bagging and boosting that explicitly create different training spaces.