Plotting Questions

[BrianMiner]

All-
I fit a model : ebm.fit(X,y). Then I am able to see overall feature importance running:

import plotly 
plotly.tools.set_credentials_file(username='XXX', api_key='123')
import plotly.plotly as plotly_py
ebm_global = ebm.explain_global()
plotly_py.iplot(ebm_global.visualize())

Question 1: Is there a way to see more than the default top features?

Then if we pass an index into visualize, we get the shape of the effect of that feature:

plotly_py.iplot(ebm_global.visualize(0))

I can see local explanations:

ebm_local = ebm.explain_local(X,y)
plotly_py.iplot(ebm_local.visualize(20))

QUESTION 2: IS it possible to get this data back and not plot it?

Question 3:

How does the above differ from show()? I have been unable to get this to work but it looks like the only difference is a drop down selector?

[interpret-ml]

Hi Brian,

Thanks for the questions! We'll work backwards through the questions as they're all related:

Question 3: How does the above differ from show()?

You're right, the only difference is the drop down -- it just makes it a little easier to find the features you're looking for. If you call show with a list (ex: show([explanation]) it will launch a dashboard to enable comparisons. What errors are you running into with the show method?

QUESTION 2: IS it possible to get this data back and not plot it?

Yes, you can. Every explanation object supports a .data() method, which returns the raw data used to generate the visualization. So in your example, you can call ebm_global.data() to get the overall feature importances, or ebm_global.data(1) to get the data for the graph in feature 1 (the same applies for ebm_local).

Question 1: Is there a way to see more than the default top features?

Right now, our built in "summary" visualization only includes the top features, but you can easily call ebm_global.data() to get the scores for every feature, then use any plotting library of your choice.

[BrianMiner]

That's fantastic!

I wont stress much about the issues with show in my environment then.

Last plotting questions, any plans in the roadmap to generate the feature shape curves with a local value highlighted (a vertical line) as in the original Ga2M paper?

[interpret-ml]

We don't have plans to support this directly in the code right now, but you can manipulate the Plotly objects with a bit of code to add it in. Here's a quick example:

from plotly.offline import iplot, init_notebook_mode
init_notebook_mode()

ebm_global = ebm.explain_global(name='EBM')
instance = X_test[:1]
feature_index = 0 # Feature to visualize


feature_plot = ebm_global.visualize(feature_index)  # Plotly object
plot_data = ebm_global.data(feature_index)   # Raw data pre-visualization 

y_min = min(plot_data['scores'])
y_max = max(plot_data['scores'])

feature_plot['layout']['shapes'] = [{
    'type': 'line',
    'x0': instance.iloc[0][feature_index],   # Value of the feature for instance 0
    'y0': y_min,
    'x1': instance.iloc[0][feature_index],
    'y1': y_max,
    'line': {    # Customize the line
        'color': 'red',
        'width': 3,
        'dash': 'dot'
    }
}]

iplot(feature_plot)

which will produce this:

You can play around with the style and shape of the line quite a bit -- here's the documentation for the plotly shapes API: https://plot.ly/python/shapes/