ehthiede
commented
6 years ago This is an issue with transform as well: .fit(x).transform(x) shouldn't be slower than .fit_transform(x). We should resolve this by adding "if Y == X: return self.dmaps" to transform so we aren't doing the expensive calculation (I just typed this though, so it probably doesn't work)
ralfbanisch
The documentation of fit_transform in diffusion_map class in incorrect:
"Fits the data and returns diffusion coordinates. equivalent to calling dmap.fit(X).transform(x)."
fit_transform is NOT equivalent to calling dmap.fit(X).transform(x). It is equivalent (returns the same output) to dmap.fit(X).transform(X), but there is computational overhead in the latter.
fit_transform(X) takes X (array-like, shape (n_query, n_features)) and returns phi (numpy array, shape (n_query, n_eigenvectors)), which represents the diffusion map eigenvectors evaluated at the data X. dmap.fit(X).transform(x) returns diffusion coordinates evaluated at the query point(s) x based on the eigenvectors computed from X.