Spectral Embedding

Spectral Embedding is on of the methods implemented in the megaman package. Spectral embedding (and diffusion maps) uses the spectrum (eigen vectors and eigen values) of a graph Laplacian estimated from the data set. There are a number of different graph Laplacians that can be used.

For more information see:

This class is used to interface with spectral embedding function. Like all embedding functions in megaman it operates using a Geometry object. The Isomap class allows you to optionally pass an exiting Geometry object, otherwise it creates one.

API of Spectral Embedding

The Spectral Embedding model, along with all the other models in megaman, have an API designed to follow in the same vein of scikit-learn API.

Consequentially, the LTSA class functions as follows

  1. At class instantiation .SpectralEmbedding() parameters are passed. See API documementation for more information. An existing Geometry object can be passed to .SpectralEmbedding(). Here is also where you have the option to use diffusion maps.
  2. The .fit() method creates a Geometry object if one was not already passed and then calculates th embedding. The eigen solver can also be passed to the .fit() function. WARNING: NOT COMPLETED Since Geometry caches important quantities (like the graph Laplacian) which do not change by selecting different eigen solvers and this can be passed and a new embedding computed without re-computing existing quantities. the .fit() function does not return anything but it does create the attribute self.embedding_ only one self.embedding_ exists at a given time. If a new embedding is computed the old one is overwritten.
  3. The .fit_transform() function calls the fit() function and returns the embedding. It does not allow for changing parameters.

See the API documentation for further information.

Example Usage

Here is an example using the function on a random data set:

import numpy as np
from megaman.geometry import Geometry
from megaman.embedding import SpectralEmbedding

X = np.random.randn(100, 10)
radius = 5
adjacency_method = 'cyflann'
adjacency_kwds = {'radius':radius} # ignore distances above this radius
affinity_method = 'gaussian'
affinity_kwds = {'radius':radius} # A = exp(-||x - y||/radius^2)
laplacian_method = 'geometric'
laplacian_kwds = {'scaling_epps':radius} # scaling ensures convergence to Laplace-Beltrami operator

geom  = Geometry(adjacency_method=adjacency_method, adjacency_kwds=adjacency_kwds,
                 affinity_method=affinity_method, affinity_kwds=affinity_kwds,
                 laplacian_method=laplacian_method, laplacian_kwds=laplacian_kwds)

spectral = SpectralEmbedding(n_components=n_components, eigen_solver='arpack',
embed_spectral = spectral.fit_transform(X)