Semi-supervised Dimension Reduction with Kernel Sliced Inverse Regression - Technologies and Applications of Artificial Intelligence - page 184

(a) PIE CMU dataset (b) COIL dataset

(c) USPS dataset (d) Adult dataset

(e) Text dataset (f) colt98 dataset

Fig. 4. Time Consumption of different datasets

a fully labeled dataset. We adopt Gaussian kernel functions, and the uniform
design(UD) model selection method[4] on 5-fold cross validation to select the
penalty parameter( c ) and Gaussian kernel width parameter( ʳ ). Each experi-
ment executes 50 times by randomly selecting the training labeled set to build
the model. We use one-tenth of the data as our testing set and the other nine-
tenths of the data as our training set. Note that as the percentage of labeled
instances in the training set changes, the entire training and testing data sets
remain the same for the entirety of that experiment.