Presented at the International Conference on Computational Science (ICCS'01)
San Francisco, CA, May, 2001.
Published in Springer Verlag's Lecture Notes in Computer Science, #2074, May 2001.

 

Model Generation of Neural Network Ensembles using Two-Level Cross-Validation

 

S.Vasupongayya, R.S.Renner, B.A.Juliano
Department of Computer Science
California State University, Chico
Chico, CA 95929-0410 USA

 

ABSTRACT
This research investigates cross-validation techniques for performing neural network ensemble generation and performance evaluation. The chosen framework is the Neural Network Ensemble Simulator (NNES). Ensembles of classifiers are generated using level-one cross-validation. Extensive modeling is performed and evaluated using level-two cross-validation. NNES 4.0 automatically generates unique data sets for each student and each ensemble within a model. The results of this study confirm that level-one cross-validation improves ensemble model generation. Results also demonstrate the value of level-two cross-validation as a mechanism for measuring the true performance of a given model.

Keywords: neural network ensembles, cross-validation, overfitting, performance measures


* Present Address: Dept. of Computer Science
California State University - Chico
Chico, CA 95929-0410
E-mail: renner@ecst.csuchico.edu