Point number limit in PCA decomposition

[hugo-pires]

Hello

Is there any limit to the number of points in PCADecomposition plot?

Thank you

[lwgray]

hi @hugo-pires Thank you for using Yellowbrick. I don’t think there is an upper-limit but i will verify this with my colleagues. Are you having problems with the PCADecomposition visualizer?

[hugo-pires]

Thank you @lwgray and congratulations for your work.

The questions is that I am projecting an aprox 5000 examples dataframe, but I'm getting an aprox 50 points. Probably they are overlaid, but I would like to doublecheck. Or use some alpha parameter.

[rebeccabilbro]

Hi @hugo-pires — thanks for checking out Yellowbrick! There probably is an upper limit for PCA, but 5k samples isn't even close to it! So I agree it's likely that the points are merely being overlaid. You can test this out using the alpha parameter, as so:

from sklearn.datasets import make_classification
from yellowbrick.features.pca import PCADecomposition

X, y = make_classification(
    n_samples=5000, n_features=200, n_informative=2, n_redundant=2
)

visualizer = PCADecomposition(alpha=.25)
visualizer.fit_transform(X, y)
visualizer.poof()

The default alpha is .75 (check out our PCADecomposition docs for more on our defaults) and ranges between 0 and 1; as it decreases, the points will become more translucent.

alpha=.75:

alpha=.25:

[lwgray]

@hugo-pires @rebeccabilbro is absolutely right that you should be able to plot 5000 points and her suggestion on altering alpha is right on. I wanted to add a few things.

1. We are actively working on the PCA visualizer and the next version will have advanced support for opacity
2. You can use proj_features=True to create a biplot which would help you interpret what's going on in the decomposition
3. Alternatively you can use a Manifold - non linear decomposition to see if it's just a linear effect
4. TSNE or isomap might tell you a lot about what's going on in the higher dimensional space

[hugo-pires]

Thank you @lwgray and @rebeccabilbro . It would also be nice to have auto colors to discrete targets, based on the levels of the label.

[lwgray]

@hugo-pires i think that #476 and #874 will address that request.