tapkee 1.3.3

Dimensionality reduction library with Isomap, LLE, t-SNE, PCA, and more

Tapkee

Efficient dimensionality reduction library with a focus on spectral methods.

Tapkee implements over 20 dimensionality reduction algorithms and provides a simple Python interface to all of them.

Installation

pip install tapkee

Usage

import numpy as np
from matplotlib import pyplot as plt
import tapkee

# Generate a Swiss roll dataset
n = 2000
t = 1.5 * np.pi * (1 + 2 * np.random.rand(n))
x, y, z = t * np.cos(t), 30 * np.random.rand(n), t * np.sin(t)
data = np.vstack([x, y, z])  # 3 features, n samples

# Unroll with Locally Linear Embedding
embedding = tapkee.embed(data, method="lle", num_neighbors=12)

plt.scatter(embedding[:, 0], embedding[:, 1], c=t, cmap="Spectral", s=5)
plt.title("Swiss roll unrolled with LLE")
plt.show()

Supported methods

Method	method=	Key parameters
Isomap	"isomap"	num_neighbors
Landmark Isomap	"l-isomap"	num_neighbors, landmark_ratio
Diffusion Maps	"dm"	gaussian_kernel_width, diffusion_map_timesteps
Laplacian Eigenmaps	"la"	num_neighbors, gaussian_kernel_width
Locally Linear Embedding	"lle"	num_neighbors
Hessian LLE	"hlle"	num_neighbors
Local Tangent Space Alignment	"ltsa"	num_neighbors
Linear LTSA	"lltsa"	num_neighbors
Locality Preserving Projections	"lpp"	num_neighbors, gaussian_kernel_width
Neighborhood Preserving Embedding	"npe"	num_neighbors
PCA	"pca"
Kernel PCA	"kpca"	gaussian_kernel_width
Factor Analysis	"fa"	max_iteration, fa_epsilon
Random Projection	"ra"
MDS	"mds"
Landmark MDS	"l-mds"	landmark_ratio
t-SNE	"t-sne"	sne_perplexity, sne_theta
Stochastic Proximity Embedding	"spe"	spe_global_strategy, spe_num_updates, spe_tolerance
Manifold Sculpting	"manifold_sculpting"	num_neighbors, squishing_rate, max_iteration

All methods accept target_dimension (default: 2). Neighbor-based methods accept num_neighbors (default: 5).

Parameters

Parameter	Type	Description
target_dimension	int	Output dimensionality
num_neighbors	int	Number of neighbors for local methods
gaussian_kernel_width	float	Width of the Gaussian kernel
landmark_ratio	float	Ratio of landmark points (0 to 1)
max_iteration	int	Maximum iterations for iterative methods
diffusion_map_timesteps	int	Number of timesteps for Diffusion Maps
sne_perplexity	float	Perplexity for t-SNE
sne_theta	float	Angle for Barnes-Hut approximation in t-SNE
squishing_rate	float	Rate for Manifold Sculpting
spe_global_strategy	bool	Use global strategy for SPE
spe_num_updates	int	Number of SPE updates
spe_tolerance	float	SPE convergence tolerance
nullspace_shift	float	Regularizer for eigenproblems
klle_shift	float	KLLE regularizer
fa_epsilon	float	Factor Analysis convergence epsilon
check_connectivity	bool	Check graph connectivity

Links

• Documentation
• Source code
• Issue tracker

Citation

If you use Tapkee in a publication, please cite:

Sergey Lisitsyn, Christian Widmer, Fernando J. Iglesias Garcia. Tapkee: An Efficient Dimension Reduction Library. Journal of Machine Learning Research, 14: 2355-2359, 2013.

License

BSD 3-Clause