| Autor: |
Widodo, Agus Wahyu, Handoyo, Samingun, Rupiwardani, Irfany, Mursityo, Yusi Tyroni, Purwanto, Imam Nurhadi, Kusdarwati, Heni |
| Předmět: |
|
| Zdroj: |
International Journal of Intelligent Engineering & Systems; 2023, Vol. 16 Issue 5, p748-759, 12p |
| Abstrakt: |
The implementation of a CNN1D model is a big challenge because it involves many hyperparameters which have to be tuned well. The research has goals to compare the classification performance of the C4.5 tree and CNN1D models applied in the imbalanced medical dataset in six performance metrics. The oversampling method is applied to the training dataset to form the oversampling training data which is divided into 5 folds of cross-validation data for tuning hyperparameters of both models. By using the grid search method, there are obtained the optimal CNN1D hyperparameters are [256, 4] for the filter and kernel sizes, and the optimal C4.5 hyperparameters are [25,7] for the minimum number of instances in the splitting node and the depth of C4.5 tree. The models with optimal hyperparameters are trained using the oversampling training data and evaluated their performance in both of original training and testing datasets. In the training dataset, the CNN1D outperforms the C4.5 tree in all six metrics with a value of 99% except in the matthews correlation coefficient (MCC) metric with 73% besides the performance gaps of both models being very large. In the testing dataset, the CNN1D also outperforms the C4.5 tree in four metrics with a value of around 99%, but in the MCC and the area under curve (AUC) metrics, the C4.5 tree outperforms the CNN1D although their performance gaps are narrow enough. An interesting future work is to use the MCC or AUC criteria in the tuning hyperparameters of models for the classification of imbalanced class datasets. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
