Possible explanation on the effect of variable selection on PAM used with SMOTE In our simulation studies with high-dimensional class-imbalanced data

In our simulation studies with high-dimensional class-imbalanced data we observed that under the null case SMOTE had hardly any effect on classification with PAM, when all the p = 1000 simulated variables where considered. On the other hand, if only a subset of the variables was used (G = 40), SMOTE seemed beneficial in reducing the class-imbalance problem of PAM, decreasing the number of samples classified in the majority class. This behavior can be seen clearly comparing the left and right panels reporting the PAM results in Figure 2.