IML and AutoML Flashcards
Deep learning neural networks, support vector machines, random forests are examples of ___ box methods.
Black
Decision trees or linear regression models are examples of ____ box methods.
White or glass
En effect in which algorithms learn undesired correlations in
the data and thus make the correct prediction for the “wrong” reason.
“Clever-Hans-Effect”
A method that can be applied to any ML algorithm.
Model-agnostic method
A method that is only applicable to a specific ML algorithm.
Model-specific method
Which are the three explainable ML models.
-Decision trees
-Linear and logistic regression models
-Nearest neighbor approach
What does LIME stand for?
- Local: only valid for the local neighborhood of an observation
- Interpretable: human understandable
- Model-Agnostic: applicable to all ML algorithms
- Explanations: provides explanations.
(True or false) In the end LIME produces a much simpler surrogate model based on the ‘local neighbors’ created
above.
True
Global agnostic method that analyzes the causal relation the features have on the prediction, then chooses a variable and fix it to a certain value for all observations and finally observes how that impacts the prediction.
Partial Dependence Plots (PDP)
Global agnostic method that measures the difference in the prediction when a variable is permuted but instead of changing a value in all observations it is restricted to local (or similar) observations.
Accumulated Local Effects (ALE)
Local agnostic method that shows the behavior of (every) observation and centers observations at zero.
Individual Conditional Expectation (ICE)
Local agnostic method that uses Shapley values as building blocks.
Shapley values
It aims to provide an end-to-end system to solve ML problems for a given training data set, automatically configure an ML model (or: pipeline) that solves the task well.
AutoML
(True or false) AutoML is rooted in HPO.
True
A high-dimensional HPO problem: Union of all algorithm’s HPs, plus a new HP that controls the chosen algorithm.
CASH (Combined Algorithm Selection and Hyperparameter Optimization)
