The dataset are so hard that even SVM can perform no better than chance (54%). ANN got 52%. On the original dataset. I thought that is impossible.