Rulefit
- Sparse linear models that include automatically detected interaction effects in the form of decision rules
- New features captures interactions between the features
- RuleFit automatically generates the interaction features from decision trees
- The decision rules are binary
- The feature importance can be calculated at the local and global level
- Interpretation of feature importance for interactions
- Bagged ensembles, random forest, adaboost can be used for generating rules
Advantages
- Automatically adds feature interactions
- Rules are easy to interpret
- For an individual only a few rules will apply
Disadvantages
- Many rules may get non-zero weight in the Lasso model
- Interpretation tricky when we have overlapping rules
Python packages
- skope-rules (Seems the development stopped 2 years ago)
- imodels
SOTA Algorithms
- Fast Interpretable greedy-tree sums (FIGS)
- Hierarchical shrinkage:post-hoc regularization of tree based methods