search

TeamHG-Memex / eli5

A library for debugging/inspecting machine learning classifiers and explaining their predictions
http://eli5.readthedocs.io
MIT License
2.76k stars 334 forks source link

Normalize coef by default #161

Open jnothman opened 7 years ago

jnothman commented 7 years ago

The predictions of a linear model are invariant to the scale of its weights. Thus the scale of weights are determined by regularisation (and, I think, the bias term if unregularised). Are weights therefore more comparable across competing linear models if scaled by default (e.g. to unit vector)?

kmike commented 7 years ago

Coefficient scale affects probability output - if coefficients are large then classifier is more "confident" - probabilities are closer to 0 and 1, at least for logistic regression.

Currently we're normalizing coefficients when computing colors in show_weights / show_prediction, so currently looking at the colors is a way to compare two models with different coefficient scale.

I can see how normalizing coefficients to unit scale (and showing this scale) can be helpful. But it also can be more confusing - what user see is no longer vanilla coefficients.