Week 7 Flashcards
Interpretability
The degree to which a human can understand the cause of a decision and consistently predict the result of a model.
Explainable
When feature values of instances can be related to model prediction in a humanly understandable way.
What levels of interpretability / explainability does Molnar state?
Interpretability is at a global level of the model, explainability is concerned with an individual prediction.
Intrinsically interpretable models
Provide all means necessary for the decisions explanation. Is interpretable due to simplicity, has all info itself.
Model agnostic explainable AI method
Explains any model, no matter the type
Model specific explainable AI method
Accesses and uses the model internals.
Counterfactual explanation
The one that is the closest to the instance that we’re trying to predict, with minimal changes, that gives a different prediction.
CNN (abbreviation)
Convolutional neural network
Expressive power
What is the structure of the explanations?
EG. is it ‘if-then’, a tree, natural language…
Translucency
Describes how much the explanation method relies on looking into the machine learning model
Portability (property of explanation method)
Measures the range of machine learning models with which the explanation can be used.
Algorithmic complexity (property of explanation method)
The computational complexity of the explanation method.
Accuracy (property of explanation)
How well does an explanation predict unseen data?
Fidelity (property of explanation)
How well does the explanation approximate the prediction of the black box model?
Certainty (property of explanation)
Does the explanation reflect the certainty of the machine learning model?
Comprehensibility (property of an explanation)
How well do humans understand the explanations? How convincing are they?
How is fidelity measured?
Objectively
How is plausability measured?
By comprehensibility: it requires a user study.
How is simulatability measured?
By measuring the degree that a human can calculate or predict the model’s outcome, given the explanation.
What should a good explanation be?
- Contrastive
- Selective
- Social
- Truthful
- General and probable
- Consistent with prior beliefs.
Black-box models
Require post-hoc explanations, cannot have perfect fidelity.
What are the two families of interpretable models that we’re focusing on?
- Linear models
- Decision trees and decision rules
Give the hierachy of linear models from big to small:
- Generalized additive models
- Generalized linear models
- Linear models
- Scoring systems
Scoring system
A specialised type of linear model that gives an integer in a range as output.