tmastny/feat

Hi there 👋

My name is Tim. I clean and analyze data, and like to build tools to make that easier. I work at Wave. I'm at home programming in R, but I also really enjoy Python.

• 🔭 I’m currently working on feat, a Python package to help with feature importances in sklearn.
• ⏭️ Up next, I'm updating my AWS sagemaker to make it easy to productionize machine learning models in R.
• 🌱 I’m currently learning the IPython API and auto-completion for siuba.
• 📖 I recently learned Hugo and updated the layout of my blog.
• 💬 Ask me about AWS and R.
• 📫 How to reach me: @timmastny on Twitter, https://timmastny.com/about/
• 😄 Pronouns: he/him

feat's Issues

feature_selection cases

Single column transformations make sense

make_pipeline(
    StandardScaler(),
    VarianceThreshold()
)

but how would this work? I don't think my logic would currently work. I'm guessing I could get a mask from the output OneHotEncoder. I think I have to think more carefully about my pipeline logic.

make_pipeline(
    OneHotEncoder(),
    VarianceThreshold()
)

Test PCA with SelectorMixin

Good reason not to nest output.

If we nested output of multi-column to multi-column transformations:

name               | feature
-----------------------------
["c1", "c2", "c2"] | ["pca1", "pca2", "pca3"]

Then it doesn't easily work with a `SelectorMixin, because we need to explode it anyway to remove one of the features. So we want

name               | feature
-----------------------------
["c1", "c2", "c2"] | "pca1"
["c1", "c2", "c2"] | "pca2"
["c1", "c2", "c2"] | "pca3"

even if it seems a little redundant. Although we could always nest at the end of the transformation.