Spectral Embedding

[aamijar]

Resolves #6486

Usage

from sklearn import datasets
from cuml.manifold import SpectralEmbedding
import cupy as cp

# (1500, 3) -> (1500, 2)
n_samples = 1500
S_points, S_color = datasets.make_s_curve(n_samples, random_state=0)

spectral = SpectralEmbedding(n_components=2, n_neighbors=None, random_state=42)
embedding = spectral.fit_transform(cp.asarray(S_points, order='C', dtype=cp.float32))


from sklearn.datasets import fetch_openml

# (70000, 784) -> (70000, 2)
mnist = fetch_openml('mnist_784', version=1)
X, y = mnist.data, mnist.target.astype(int)

spectral = SpectralEmbedding(n_components=2, n_neighbors=None, random_state=42)
embedding = spectral.fit_transform(cp.asarray(X, order='C', dtype=cp.float32))

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[cjnolet]

Please pull these primitives into RAFT. We centralize these primitives specifically to 1) reduce code duplication, and 2) highlight areas that could be further optimized in the future- optimizations of these types of computations propagate to all algorithms which use them. Having them buried in each algorithm reduces the maintainability as well as the benefits from continued optimization.

[cjnolet]

This should be a preprocessing method in cuVS c++ API (since it's nearest neighbors + laplacian + eigendecomp). It should be exposed in cuML. We should also be allowing other options for the nearest neighbors than just brute-force.

[cjnolet]

Just a note for future use as we definitely want to make sure we are using the new raft sparse types (raft::sparse::device_coo_matrix) in public APIs (this is not a public API). We are trying to get rid of the legacy types (raft::sparse::COO).

[cjnolet]

See comments in review