MostInformativeFeatures visualizer

[majidaldo]

Is your feature request related to a problem? Please describe. This feature is motivated by the discussion in #510. The problem is how to visualize feature importance for multiple classes and/or instances. This requires solving two problems: 1. selecting the "most informative features" and 2. producing an appropriate visualization.

Describe the solution you'd like MIFV = MostInformativeFeatures MIFV().predict(X) would give a viz for the features most responsible for the prediction MIFV().poof()/fit(X) would show a viz for the features

For both cases, the proposed visual is a heat map where one axis is class labels or data and the other axis represents features.

Questions/Issues:

• The estimators would have to (mathematically) support such a notion of feature importance.
• Top 10 features doesn't make sense if all the features are similar in strength.
• Does it make sense to numerically compare feature strengths over multiple classes?

[ndanielsen]

@majidaldo thanks for creating this issue and following / linking with other discussions.

I look forward to the community thinking through the questions and issues on this. =)

[rebeccabilbro]

Going to go ahead and close this one out since it's gone a bit stale; to anyone with interest/bandwidth — feel free to reopen!

[bbengfort]

@rebeccabilbro this is one that is very interesting to me, perhaps I will take a crack at it later in the semester.

[rebeccabilbro]

That would be great @bbengfort, reopening!