Figure 7: Leave-one-out cross-validation suggests that the individual cluster assignments are robust and specific.
From: A genetic and computational approach to structurally classify neuronal types
(a) Schematic of the leave-one-out study. The best match cluster is the newly formed cluster whose cells are structurally most similar to the deleted cell. (b) Distribution of the similarity index over the whole data set. The similarity index is a measure of prospective cell type assignment capability; 0 denoting the worst score and 1 denoting the perfect score (see text). (c) Distribution of the Rand index over the whole data set, showing that reassignment of cell types rarely change after a cell is deleted. Insets: binarized matching matrices generating the lowest and the highest Rand indices, where each row (column) corresponds to a cluster of the original (one-left-out) clustering. A white pixel indicates a nonzero intersection and a black pixel indicates a disjoint cluster pair. (d) Distribution of the discovered number of clusters over the whole data set, demonstrating the effect of taking one cell out on the resulting number of clusters. Reclustering the data set without the left-out cell virtually always resulted in 15 clusters. The y axis is common to (b–d).
