Human-Centered Interpretable Machine Learning
•
3 likes•588 views
Why should we care? How to explain a complex model? Use case FIFA 2019 Use case Credit Scoring When to use XAI? Make models not wars
Recommended
More Related Content
Similar to Human-Centered Interpretable Machine Learning
Similar to Human-Centered Interpretable Machine Learning (20)
Recently uploaded
Recently uploaded (20)
Human-Centered Interpretable Machine Learning
- 1. Explore, Explain, and Debug Human-Centered Interpretable Machine Learning Przemysław Biecek
- 2. Model explanations Why should we care?
- 6. • “You don’t see a lot of skepticism,” she says. “The algorithms are like shiny new toys that we can’t resist using. We trust them so much that we project meaning on to them.” • Ultimately algorithms, according to O’Neil, reinforce discrimination and widen inequality, “using people’s fear and trust of mathematics to prevent them from asking questions”. https://www.theguardian.com/books/2016/oct/27/cathy-oneil-weapons-of-math- destruction-algorithms-big-data 6 Cathy O'Neil: The era of blind faith in big data must end black boxes Why do we need explanations for complex models?
- 8. 8
- 9. How to explain a complex model?
- 13. Source: https://www.kaggle.com/karangadiya/fifa19/data library("gbm") fifa_gbm <- gbm(Value~., data = fifa19, n.trees = 250, interaction.depth = 3)
- 14. Full name Ádám Csaba Szalai Date of birth 9 December 1987 Place of birth Budapest, Hungary Playing position Striker wikipedia Prediction from GBM model: 7 344 354 EUR How? Why? Really?
- 15. Full name Ádám Csaba Szalai Date of birth 9 December 1987 Place of birth Budapest, Hungary Playing position Striker wikipedia
- 16. Full name Péter Gulácsi Date of birth 6 May 1990 Place of birth Budapest, Hungary Playing position Goalkeeper wikipedia
- 21. Performance for selected modeling methods red models are the most interesting https://buecker.ms/slides/crccxvi/slides.html#1
- 23. Partial Dependency Plot for the most important feature https://buecker.ms/slides/crccxvi/slides.html#1
- 24. Single customer
- 27. The most common questions for XAI
- 28. https://kmichael08.github.io XAI-BOT: What would you ask for?
- 29. https://kmichael08.github.io What If? Why? How good is the prediction? XAI-BOT: What would you ask for?
- 30. When to use XAI?
- 31. Domain understanding Predictive modeling Validation and Justification EDA, transformations Linear models MSE, p-values Shift in our focus: Statistics
- 32. Domain understanding Predictive modeling Validation and Justification Domain understanding Predictive modeling Validation and Justification EDA, transformations Linear models MSE, p-values Simple EDA Lots of models + optimisation Test/train Cross Validation Shift in our focus: Machine Learning
- 33. Domain understanding Predictive modeling Validation and Justification Domain understanding Predictive modeling Validation and Justification Domain understanding Predictive modeling Validation and Justification EDA, transformations Linear models MSE, p-values Simple EDA Test/train Cross Validation Simple EDA AutoML XAI, Fairness, Ethics Validation and Justification Shift in our focus: Human Oriented ML? Lots of models + optimisation
- 34. MDP : : Model Development Process Data validation Feature selection Parameters tuning Problem formulation Crisp modelling Fine tuning Maintaining Data acquisition Model deployment Data cleaning Data exploration Sample selection Feature engineering Model selection Model validation Documentation Communication Data preparation Data understanding Model delivery Model assembly Model audit Model benchmarking Iterations P1 C1 C2 F1 F2 M1 M2 M3 time MDP - process for model development
- 35. MDP : : Model Development Process Data validation Feature selection Parameters tuning Problem formulation Crisp modelling Fine tuning Maintaining Data acquisition Model deployment Data cleaning Data exploration Sample selection Feature engineering Model selection Model validation Documentation Communication Data preparation Data understanding Model delivery Model assembly Model audit Model benchmarking Iterations P1 C1 C2 F1 F2 M1 M2 M3 time MDP - process for model development
- 36. MDP : : Model Development Process Data validation Feature selection Parameters tuning Problem formulation Crisp modelling Fine tuning Maintaining Data acquisition Model deployment Data cleaning Data exploration Sample selection Feature engineering Model selection Model validation Documentation Communication Data preparation Data understanding Model delivery Model assembly Model audit Model benchmarking Iterations P1 C1 C2 F1 F2 M1 M2 M3 time MDP - process for model development
- 37. MDP : : Model Development Process Data validation Feature selection Parameters tuning Problem formulation Crisp modelling Fine tuning Maintaining Data acquisition Model deployment Data cleaning Data exploration Sample selection Feature engineering Model selection Model validation Documentation Communication Data preparation Data understanding Model delivery Model assembly Model audit Model benchmarking Iterations P1 C1 C2 F1 F2 M1 M2 M3 time MDP - process for model development
- 38. MDP : : Model Development Process Data validation Feature selection Parameters tuning Problem formulation Crisp modelling Fine tuning Maintaining Data acquisition Model deployment Data cleaning Data exploration Sample selection Feature engineering Model selection Model validation Documentation Communication Data preparation Data understanding Model delivery Model assembly Model audit Model benchmarking Iterations P1 C1 C2 F1 F2 M1 M2 M3 time pre-modelling explore in-modelling explain post-modelling debug pre-, in-, post- modelling explainability
- 39. Model specific / agnostic explainability
- 40. Model specific Model agnostic + Exploits the structure of a model + Allow for deep diagnostic - Explanations cannot be easily compared between models - You need to create an explainer for every possible model structure Examples: • randomForestExplainer • EIX - lightgbm/xgboost explainer + Explanations can be compared between different models + Can be used to any model (model is a black box) - Explanations are (im most cases) approximations, they may be inaccurate or wrong - It is easy to miss some quirks Examples: LIME / SHAP / Break Down / Partial Dependency Profiles / Permutational Feature Importance
- 42. Local explanations Global explanations + Focused on a single observation + Good for dissecting of a single prediction + Good for debugging - Feature effects may be different for different observations - Tricky for correlated features Examples: • LIME / SHAP / Break Down • Ceteris Paribus + Focused on a model + Good for global model summaries + Good for comparisons - Local behaviour may be different - For non-additive models they may be to simplistic Examples: • Feature Importance • Partial Dependency Profiles
- 43. An Introduction to Machine Learning Interpretability Navdeep Gill, Patrick Hall https://www.h2o.ai/oreilly-mli-booklet-2019/ Interpretable Machine Learning Christoph Molnar https://christophm.github.io/interpretable-ml-book/ Predictive Models: Explore, Explain, and Debug Przemyslaw Biecek and Tomasz Burzykowski https://pbiecek.github.io/PM_VEE/
- 46. model explainer https://medium.com/@ModelOriented/xai-in-the-jungle-of-competing-frameworks-for- machine-learning-fa6e96a99644 DALEX wraps a model available in R or python or through REST API into a standardised container. Other tools for XAI may then work on the model despite its internal structure.
- 48. Thank you!