tpot2
tpot2 copied to clipboard
Genetic Feature Selection Improvements
Feature selection can be done using the FeatureSetSelector by specifying groups of columns as a "set" or by simply selecting a single column per node. Each set or feature will be encoded in an individual FeatureSetSelector node. To mix and match sets/features, the algorithm must manually move around these nodes.
This can be inefficient with a high number of individual columns or sets to select from. This is because the mutation and crossover operations would generally move single nodes around. Multiple nodes may be included in a crossover, but only if they come attached to a selected downstream node.
Another option could be to create a special case node within the GraphIndividual that has its own mutation and crossover functions. These mutation/crossover functions would be able to mix and match a large number of features in a single operation, similar to other implementations of genetic feature selection in other packages.