Isomap — distances measured along the cloud
Isomap connects each row to its nearest neighbors, computes shortest paths on that graph, and embeds the rows so Euclidean distance in the new space matches those path lengths. The path length is a geodesic: it follows the cloud instead of cutting through empty space. That is what unfolds a Swiss roll, where the straight-line distance between two points jumps off the roll and the meaningful distance walks along it.
The neighbor count has to be large enough that the graph has one connected component. A disconnected graph has infinite distances and the embedding is meaningless. It also has to be small enough that shortcuts do not jump from one part of the manifold to another. When a shortcut appears, the unfolding collapses.
On a table that really is a filled blob, there is no manifold to unfold and Isomap spends its work approximating ordinary distances. PCA is the honest tool in that case.
Settings that change this fit
- n neighbors
- Edges per row in the geodesic graph. Raise it until the graph is connected, and stop before edges leap across gaps.
- n components
- Dimension of the embedding that tries to match the path lengths. Two for a picture, more if a later model will use the coordinates.