ENH Emphasize discussion on multi-class classification in tree notebook #730
Conversation
There was a problem hiding this comment.
I pushed some commits to adjust the colormaps to:
- avoid using a binary Red Blue colormap for multiclass decision boundary,
- use a continuous viridis for the predict_proba plots because 0.5 is not meaningful any more in a one vs rest setting.
Below are some more suggestions, but beyond this, LGTM.
python_scripts/trees_sol_01.py
Outdated
| # For example, in the plot below, the first plot on the left shows in red the | ||
| # certainty on classifying a data point as belonging to the "Adelie" class. In | ||
| # the same plot, the blue color represents the certainty of **not** belonging to | ||
| # the "Adelie" class. The same logic applies to the other plots in the figure. |
There was a problem hiding this comment.
@ogrisel does this paragraph make sense to you when using a diverging colormap?
Or can you please elaborate on how the 0.5 probability cannot be interpreted under this one-vs-rest logic?
There was a problem hiding this comment.
Indeed this paragraph needs to be updated with the new colors (e.g. bright yellow vs dark purple). What I mean is that the chance level for a one vs rest binary classification problem that comes from a multi-class classification problem is almost never at 0.5. So using a colormap with a neutral white at 0.5 might give a false impression.
When we do one-vs-rest, we do not threshold the value of predict_proba at 0.5 to get the hard class predictions but instead concatenate of the 3 one-vs-rest predict_proba vectors into a 2D array and take the argmax across the classes dimension.
There was a problem hiding this comment.
Does it make sense to keep the colorbar at the bottom of the plot in this case?
Co-authored-by: Olivier Grisel <[email protected]>
Co-authored-by: Olivier Grisel <[email protected]>
Co-authored-by: Olivier Grisel <[email protected]>
…ree notebook (#730) Co-authored-by: ArturoAmorQ <[email protected]> Co-authored-by: Olivier Grisel <[email protected]> c9a7ad4
This PR reworks the general wording to avoid redundant text and prefer verbs in present mode.
It also adds a plot of
predict_probaper class, as inspired by the Plot classification probability example