Using a confusion matrix to measure accuracy in multiclass problems