search

viadee / javaAnchorExplainer

Explains machine learning models fast using the Anchor algorithm originally proposed by marcotcr in 2018
BSD 3-Clause "New" or "Revised" License
15 stars 3 forks source link

Ignoring features #19

Open TobiasGoerke opened 4 years ago

TobiasGoerke commented 4 years ago

The complexity of constructing anchors mainly depends on the number of features. Thus, if an instance has many features, the algorithm is slow. The parameter maxAnchorSize allows to constrain the rule so that it stops if a certain number of predicates are added. However, the alogithm does not allow to completely ignore a feature in a way that no rule gets created containing ignored features. We should think about adding this feature. In some cases, the users knows or suspects which features are important (e.g., by previously obtaining feature importances from a RF) and does not want the overhead of taking all features into consideration.

TobiasGoerke commented 4 years ago

The algorithm would still have to perturb ignored features to mitigate feature interaction effects.