Add multinomial version of credit card notebook (monotonic XGBoost, PDP, ICE, & Shapley) about interpretable-ml HOT 5 CLOSED

My advice here is ... (see what you think)

1. GENERAL

• scrap monotonic (since it is broken/unsupported).

2. GLOBAL VAR. IMP.

• do not use XGBoost var. imp. (it's theoretically inconsistent).
• use mean absolute shapley across all classes and rows as overall var. imp. ranking then just make a horizontal bar chart of the above values for each var. for overall var. imp. across all classes. (Should look like XGBoost var. imp. chart.)
• Then break shapley out by class using shap package summary plot, like you did for PAY_0 = 0, but just repeat for all classes.

3. LOCAL VAR. IMP. (show for 2-3 rows)

• show mean absolute shapley ranking across classes for each variable as horizontal bar chart for a single row. (should look like global var. imp. horizontal bar chart above.)
• show regular signed shapley values and reason codes for the class that the model predicts.

4. PDP/ICE

• Keep overall PDP plot you have, but make a point that if classes > 10, we would only take the top 10 most common classes by default.
• For ICE, plot at min, max, and deciles as we've done before, but now we have to do it for each class. Also keep partial dependence in ICE plot and plot ice curves in a color gradient of the same color as the corresponding partial dependence. So ICE for class=0 would be all blue gradients, for class=1 it would be all orange color gradients, and so on.

from interpretable-ml.

thx @jphall663 !

Here are my responses:

1. • Scrap monotonic -> Yes
2. • Will remove XGBoost global var importance
   • Will do mean absolute for Shapley across all classes and rows and make a horizontal bar chart.
   • Will break Shapley out by class (Already thought of doing this but wanted to get a simple notebook out first)
3. • Will repeat (3) and (4) for a single row. is that what you mean?
4. • PDP/ICE comments seem fine to me.

from interpretable-ml.

For local var. imp.:

• do like we originally did for multinomial loco, just mean absolute shapley value across all classes for each variable.
• then make reason codes with shapley the way we have been doing, but only for the predicted class. not all classes.

from interpretable-ml.

Moved DT surrogate comments here: #5

from interpretable-ml.

I think I have addressed all of your comments from a code standpoint.

from interpretable-ml.