In recent years, there has been a surge of interest in spectral
manifold learning techniques. Despite the interest, only little work
has focused on the empirical behavior of these techniques. We construct
synthetic data of variable complexity and observe the performance of
the techniques as they are subjected to increasingly difficult problems.
We evaluate performance in terms of both a classification and a regression
task. Our study includes Isomap, LLE, Laplacian eigenmaps, and
diffusion maps. Among others, our results indicate that the techniques
are highly dependent on data density, sensitive to scaling, and greatly
influenced by intrinsic dimensionality.