The IFS curves for the 3 datasets.

<p>A series of classifiers can be constructed using different number of top features from the mRMR tables during the IFS process. Plot showing the performances of the different classifiers, with <i>MCC</i> as the main measurement on the y-axis. As the classifiers used different numbers of features, we represented the classifiers with the corresponding number of features used in x-axis. In Dataset 1, the highest <i>MCC</i> (0.7046) was achieved at 118 features. This finding demonstrated that the classifier adopting the top 118 features in the mRMR table for Dataset 1 performed the best, and the 118 features were regarded as composing the optimal feature set for Dataset 1. Similarly, a peak of <i>MCC</i> 0.7322 and 0.7267 was obtained at 35 and 41 features in Dataset 2 and Dataset 3, respectively. These results demonstrated that by using the top 35 and 41 features in the mRMR table for Dataset 2 and Dataset 3, respectively, the classifier performed the best. The 35 and 41 features were regarded as composing the optimal feature set for Dataset 2 and Dataset 3, respectively.