THE EFFECT OF RATIO BETWEEN TRAINING AND TESTING SET IN MODEL SELECTION FOR NEURAL NETWORKS CLASSIFICATION
| Main Authors: | Rezeki, Sri, Subanar, Subanar, Suryo , Guritno |
|---|---|
| Format: | Article PeerReviewed application/pdf |
| Bahasa: | eng |
| Terbitan: |
North Texas University
, 2006
|
| Subjects: | |
| Online Access: |
https://repository.ugm.ac.id/32931/1/1.pdf https://repository.ugm.ac.id/32931/ |
Daftar Isi:
- Determining the l arge of sample size is highly context dependent. In ge neral, the larger the dimension of parameter, the larger sample size must be to obtain a given degree of approx imation. In m;my cases of NN application, the data set is randomly split into two muwally exclusive sub!> ts. i.e. training and testing sets. The first is used for model building, while the second is used t o assess the performance (generalization) of the model. Both training •and test ing sets are the sa me size. In this paper, five dif'lcrc n t datr1 partitioning is uti li zed to test whether the prediction abi lity of NN is affected by the number of observation in training set. The purpose of this paper is to evaluate the eiTcct o f ratio between training and testing sets to the misclassitication rate in neura l networks (NN) model. An empirical study has been done by using Fisher's iris da t a. The results show that the misclassifieation rate decrease when the number of training set increase. Model with 2 hidden neuron obtains minimum error when the ratio of train ing set is 20%. Whereas model with I h idden neu ron yield" the m!!limum error at the ratio of training set 40%, 50%, 60% and 80%