Roc plots with thresholds clearly identified. Using Jupyter ipywidgets one could also think of making it interactive where the user can specify the threshold based on which the TPR/FPR might change. Something similar to what h20-3 does.

Follow same outline here but for multinomial prediction: https://github.com/navdeep-G/interpretable-ml/blob/master/notebooks/credit/xgb_credit_binary_classifier.ipynb

• Need multinomial XGBoost model with monotonicity constraints
• Need variable importance from model
• Need Shapley values per class outcome
• Need global Shapley variable importance
• Need local Shapley example
• PDP (partial dependence per class outcome)
• ICE (ICE values per class outcome)

• https://github.com/navdeep-G/interpretable-ml/blob/master/notebooks/simulated/binomial/xgb_simulated_binomial_shap_pdp_ice.ipynb

• Fix repetitive info in notebooks
• Fix comments

Anything else?

Thinking of ways to have comparative chart for comparing different methods and accuracy.
Checking some old papers might help on how to summarize results from different algorithms.

• DIA
• LIME
• DT Surrogate and LOCO
• Residual and sensitivity analysis

If time permits, it might be a good idea to normalize the shapely scores such that it is bounded between 0 and 1 so that we get distinct color when the influence of a feature is positive(red)/negative(blue).
Current bound is -1, 1 and the color map is absolute so red and blue get mixed and one has to focus on the color legend to understand what's going on.

• In order of importance:
• Ensure it works for all cases: binomial, multinomial,and regression
• Eventually should be part of a package
• Add tests

• Abstract
• Introduction
• Simulated data
  • Global analysis
    • Surrogate DT
    • Decision boundary plot
    • Comparison of global var. imp. methods
  • Local analysis: Comparison of local var. imp. methods
• Credit card data
  • Global analysis: Surrogate DT, decision boundary, global shap
  • Local analysis: Local shap
• Conclusion

• Utilize datatable for building confusion matrices and conducting DIA analysis

• Adverse Impact: (tp + fp) / (tp + fp + tn + fn)

• Adverse Impact Disparity (Ratio): non-reference adverse impact / reference adverse impact

• Adverse Impact Parity: low_threshold < Adverse Impact Disparity < high_threshold

This is a low hanging fruit which could be covered under Exploratory Analysis.
We can then compare the high correlated features and not highly correlated features to understand or poke around Surrogate Models validate if the influences are rightly captured.

If time permits, it might be worth to have a CV plot highlighting the sweet spot during the model training phase.

So far we show 1-d plot, it might we worth showing 2-way interactive plots as well.
I will add this to the simulated notebook. It will complement 1-way PDP and ICE.

• Healthcare
• Insurance
• Marketing
• Sports

• Binomial case
• Multinomial case
  • Find some package, even if it's R, that does multinomial decision tree for one overall surrogate DT.
  • Be sure to show overall error and CV error.
  • From @jphall663: Please try to find a single, multinomial DT. I promise this is the most elegant way to handle. It will show how to get to probabilities in each class, but in a single tree. We have to summarize info to be useful/novel. Per class surrogate DT is ok too, but I would only show those after showing a single DT, like this one:
    
• Regression case

Would suggest having a confusion matrix plot, I think we already have access to the values.

1. Values could be normalized to be represented as %ages
2. or not normalized as well. Using the computed values to draw the plot.
   if time permits one could think of some cool intuitive visualization as well.