Explainable ML

[choshin84]

Tweet summary

PDP (Partial Deference Plot)

• Assumption: Fix all other variables, generate simulation on target variables
• Biggest issue: non-realistic parameter combination, i.e. 200cm 40kg --> ALE: only within small section, calculate prediction diff. Cons: No equivalent to ICE Shaky result based on section size
• Better to use with ICE to plot each sample simulation

Global surrogate

• Interpretable ML to fit ML which fit originally to the data
• Use original data set and ML prediction as output

Local surrogate (LIME)

• Each prediction interpretation
• Use generated data “nearby” and ML prediction as output
• Nearby definition is fuzzy and result stability is problematic Shapley value
• Local surrogate with game theory
• Calculate all possible feature combination with 1 fixed feature value, avg is shapely value
• More full explanation as it considers all possible combinations for the 1 feature value
• Lots of computing time SHAP
• Computationally light version of Shapley, and combine LIME linear additive model for features
• Feature interaction can be also visualized

Useful link

https://christophm.github.io/interpretable-ml-book/agnostic.html

[choshin84]

[choshin84]

not model-based (prediction based) partial dependence curve approach https://arxiv.org/pdf/1907.06698.pdf

[choshin84]

https://github.com/parrt/stratx