Figure 1: Schematic of the margin classifier.

Figure 2: Schematic of discriminative margin clustering.

Figure 3: Heatmap for the toy example.

Figure 4: Dendograms for hierarchical and discriminative margin clustering on the toy example.

Figure 5: Training and test errors for the toy example.

Figure 6: Dendogram for the lung samples. Note that there is no common feature set for the large cell sub-type, though traditional hierarchical clustering groups these samples together. Also, the sub-groups for the other classes are different from those produced by traditional clustering methods.

Figure 7: Margin for the test samples along with the label of the cluster which yields the largest margin. The vertical lines separate the actual classes of the test samples, while the color of the points illustrate the predicted class.

Figure 8: Training and test errors for the samples.

Figure 9: Heatmap for the expanded gene-list of a breast cancer sub-type. Note the spread of expression values in the normal samples. Also note that ERBB2, GRB7 and LIV-1 define one sub-class, while ESR1 and NAT1 define the other sub-class.