1. Active learning algorithm require robust classifier for robust usefulness estimation. Therefore, we choose generative classifiers, namely the Parzen window classifier (PWC)
题外
As the optimization of the overall performance level is not the scope of this paper, we choose to simply standardize each attribute (Z-standardization) and use an univariate Gaussian kernel with fixed standard deviation of sigmma=0.7 for all datasets and active learning algorithm. This ensures fair comparability that is independent of a classifier bias.