RATE meeting
July 2nd
Previous meeting
Formalization
A Machine Learning explanation \(E\) is an answer to the question:
Why are (instances of) \(\mathcal{X}_E\) classified as \(y\)?
where \(\mathcal{X}_E \subseteq \mathcal{X}\) and \(y \in \mathcal{Y}\).
Philosophical models
(Deductive-Nomological, Inductive-Statistical, Statistical-Relevance models)
Conclusion
Put on the stack, work towards a practical application.
Question: how to formally define an explanation?
Overview
Question: how to visualize the trade-off between explanation desiderata?
- Triangle
- Metrics
- Two alternatives
- RATE meeting
Triangle
Metrics
Generic: \(\frac{|(y \in y_{expl}\ |\ y = NULL)|}{|y_{expl}|} \)
with \(y_{expl}\) the predicted classes by the explanation.
Simple: \(1 - \frac{N_{expl}}{N_{ref}} \)
with \(N\) the number of non-leaf nodes in the decision tree.
Accurate: \(\frac{\Sigma TP\ +\ \Sigma TN}{|(y \in y_{expl}\ |\ y \neq NULL)|} \)
but normalized per class. Only explained instances!
Problem
Deck July 2nd
By iamdecode
