I'm trying to do a multiclass classification with h2o in R. I stacked a model with a RF, GBM and deeplearning. The accuracy is ok (~0.81), but the average F1 score is bad because class B has a very high error rate. I understand 1 reason is because class B has very little samples (same for training set). Theoretically, what would be a good way to improve the accuracy/lower the error rate for class B?
A Error Rate: 0.1027 = 467 / 4,547 B Error Rate: 0.6847 = 393 / 574 C Error Rate: 0.2347 = 721 / 3,072
Total Error Rate 0.1930 = 1,581 / 8,193