Feature sets/groups

[strukevych]

Some features make sense to be dropped out only together.

It would also be a good feature to try different feature sets and how they can learn.

For example, on my dataset, I have a lot of features calculated out of 4 main features.

So all the features strongly correlated to these 4 main features
So without this feature, I'm achieving 60% accuracy as max, and without them is 67% max because of the spread of feature importance
But, without calculated features and only with these 4 main features, I'm achieving 65% max.

So they are working badly only together

To get the best model I need to run automl two times and instead of that we could use feature sets and feature groups