Due to the time limitation, I did not implement the measurement metrics of MLP. Furthermore, the testing on MLP is extremely complex and time consuming due to the various parameters involved. For example, all those parameters, hidden neurons in the hidden layer, number of layers, learning rate, and number of epochs, would affect the test result significantly. To minimized testing work load, I just compared MLP with other two remarkable classifiers, Random Forest and SVM, and used grid search algorithm to find the structures and parameters of learning performance(Table 2) with 1000 features selected(K).
The following parameters of MLP were examined: hidden neurons in the hidden layer={100 to 500}, number of layers={1,2,3}, learning rate={0.05,0.1,0,2,0.3}, and number of epochs={10000 to 50000}.
| Method | Parameters of the Best Performance | Acc. [%] |
|---|---|---|
| MLP | # of neurons\(=\) 300, #of layers\(=\) 1, learning rate\(=\) 0.1, # of epochs\(=\) 40000 | 82.8 |
| SVM | Kernal Function\(=\) linear | 80.2 |
| Random Forest | # of trees\(=\) # of Selected Features(\(K\)) | 79.4 |