Accepted in Journal of Big Data
Available package: pip install alpha-pca
| 1 to 80 Components | 16 Components | 64 Components |
|---|---|---|
 |
 |
 |
Alpha-PCA is more robust to outliers than standard PCA.
Standard PCA is a special case of alpha PCA (when alpha=1).
The model is inherited from a sklearn module and works the same way as the standard PCA.
It also supports PyTorch tensors (on cpu and GPU).
from alpha_pca import *
import torch
X = torch.randn(16, 10) # also works with numpy
pca = AlphaPCA(n_components=5, alpha=0.7, random_state=123) # alpha=1 -> standard PCA
pca.fit(X)
# to project X in the latent space
X_transformed = pca.transform(X) # (16, 10) -> (16, 5)
# fit inverse
X_ = pca.inverse_transform(X_transformed) # (16, 5) -> (16, 10)
# directly approximate X_ == inverse_transform(transform(X))
X_ = pca.approximate(X) # (16, 10) -> (16, 10)
# Find the optimal alpha via a reconstruction loss
best_alpha = pca.compute_optimal_alpha(X, n_components=5)Experiments are located in the experiments/ folder.