The Five Tribes of Machine Learning Explainers
The document summarizes five "tribes" or categories of machine learning explainers: 1) The Featurists, who identify important features that a model relies on through methods like feature importance, selection, and correlation. 2) The Speculators, who examine how a model responds to changes in individual variables using techniques like partial dependence plots and individual conditional expectations. 3) The Localizers, who fit interpretable models locally to explain individual predictions using methods like LIME and anchors. 4) The Convoluters, who visualize important regions in images for convolutional neural networks. 5) The Trainalyzers, who identify training examples that most influenced a prediction using influence functions.
Convert to study materialsBETA
Transform any presentation into ready-made study material！select from outputs like summaries, definitions, and practice questions.
The Five Tribes of Machine Learning Explainers
- 1. The Five Tribes of ML Explainers (and what you can learn from each) Micha? ?opuszy┰ski PyData Berlin, 07.07.2018
- 2. Intro Looking at performance metrics for your model is great but....
- 3. Intro Looking at performance metrics for your model is great but.... Why? Why? Why? Why? Why? Why? your model, predicts what it predicts?
- 5. The Featurists Feature importance? Feature importance from models (RF, ExtraTrees, Boosted trees, linear models)? Model agnostic feature importance (e.g. permutation importance from ELI5)? Feature selection? Filters - e.g. filtering most correlated features? Wrappers - e.g. forward/backward selection? Embedded methods - e.g. lasso? Idea: Find features important to your model? all my features Model Class Reliance: Variable Importance Measures for any Machine Learning Model Class, from the ＾Rashomon￣ Perspective (2018)
- 7. The Speculators Idea: Check how your model responds to change of one variable? Answer 1: Partial dependence plots:? Example from free book Interpretable ML by Christoph Molnar
- 8. The Speculators Idea: Check how your model responds to change of one variable? Answer 2: Individual Conditional Expectations? Example from free book Interpretable ML by Christoph Molnar
- 10. The Localizers Idea: Fit interpretable model, which is locally correct? Simple model = Linear Why Should I Trust You? Explaining the Predictions of Any Classifier Ribeiro, Singh, Guestrin LIME Simple model = Rules Anchors: High-Precision Model-Agnostic Explanations Ribeiro, Singh, Guestrin Anchors
- 11. The Convoluters
- 12. The Convoluters (only for ConvNets) Two Ideas: Visualize the important regions in the image? Example: Labrador One of the important high level features for Labradors + high level features "Labradorish" parts of image Explanation Figures from The Building Blocks of Interpretability, Olah et al, distill.pub
- 13. The Trainalyzers
- 14. The Trainalyzers Idea: Which training examples contributed mostly to a given prediction? Understanding Black-box Predictions via Influence Functions, W. Koh, P. Liang Sample approach - influence functions
- 15. I collect links to interesting papers & soft @lopusz github.com/lopusz/awesome-interpretable-machine-learning There is a lot more!