108
Z. Liu et al.
useful information carried is highly redundant, while the data noise is large. This
will result in a low signal-to-noise ratio and poor prediction performance. When the
number of feature variables is too small, which results in too little useful information
carried and it causes poor prediction performance. Therefore, too many and too
few feature variables are not conducive to the prediction performance of the model.
Only when the appropriate number of feature variables are combined, the MLP can
perform best.
Comparison of Prediction Performance
Two groups of comparative experiments are designed, the first group is a comparison
between SVR and CART-RFE-SVR, and the second group is a comparison between
MLP and CART-RFE-MLP. CART- and CART-RFE-MLP use SVR and MLP as the
feature subspace prediction models mentioned in this article, respectively. As shown
in Fig. 7, the predicted value and prediction error of 30 selected test samples are
compared. The predicted value and prediction absolute error of SVR and CART-RFESVR are compared, respectively, in the upper and lower layers of (a), and CART-RFESVR predicted value is closer to the true value and its errors are generally smaller.
The predicted value and prediction absolute error of MLP and CART-RFE-MLP
are also compared, respectively, in the upper and lower layers of (b), and CARTRFE- MLP predicted value is closer to the true value and its errors are generally
smaller. Therefore, (a) and (b) intuitively show that the prediction performance of
CART-RFE-SVR and CART-RFE-MLP is better than SVR and MLP, respectively.
As shown in Table 5, there are two groups of comparisons of the model’s prediction performance on the test set. The bold fonts in the table indicate better indicators
in every group of comparison. Compared with SVR, CART-RFE-SVR is better than
SVR in four indicators. Compared with MLP, CART-RFE-MLP is superior to MLP in
three indicators, and only slightly inferior to MLP in ACC. Two groups of comparative experiments show that the integrated learning method in this paper has improved
prediction performance compared with a single model.
Fig. 7 Comparison of predicted value and prediction error. (Color figure online)
Z. Liu et al.
useful information carried is highly redundant, while the data noise is large. This
will result in a low signal-to-noise ratio and poor prediction performance. When the
number of feature variables is too small, which results in too little useful information
carried and it causes poor prediction performance. Therefore, too many and too
few feature variables are not conducive to the prediction performance of the model.
Only when the appropriate number of feature variables are combined, the MLP can
perform best.
Comparison of Prediction Performance
Two groups of comparative experiments are designed, the first group is a comparison
between SVR and CART-RFE-SVR, and the second group is a comparison between
MLP and CART-RFE-MLP. CART- and CART-RFE-MLP use SVR and MLP as the
feature subspace prediction models mentioned in this article, respectively. As shown
in Fig. 7, the predicted value and prediction error of 30 selected test samples are
compared. The predicted value and prediction absolute error of SVR and CART-RFESVR are compared, respectively, in the upper and lower layers of (a), and CART-RFESVR predicted value is closer to the true value and its errors are generally smaller.
The predicted value and prediction absolute error of MLP and CART-RFE-MLP
are also compared, respectively, in the upper and lower layers of (b), and CARTRFE- MLP predicted value is closer to the true value and its errors are generally
smaller. Therefore, (a) and (b) intuitively show that the prediction performance of
CART-RFE-SVR and CART-RFE-MLP is better than SVR and MLP, respectively.
As shown in Table 5, there are two groups of comparisons of the model’s prediction performance on the test set. The bold fonts in the table indicate better indicators
in every group of comparison. Compared with SVR, CART-RFE-SVR is better than
SVR in four indicators. Compared with MLP, CART-RFE-MLP is superior to MLP in
three indicators, and only slightly inferior to MLP in ACC. Two groups of comparative experiments show that the integrated learning method in this paper has improved
prediction performance compared with a single model.
Fig. 7 Comparison of predicted value and prediction error. (Color figure online)