As Pimaster advises, remove/replace features that would typically increase the turnover in your traditional portfolio, especially if you have a target that is short.
Don't go wild with your hyperparameters. I find "lightgbm III" to be a good base; don't divert too much from it and add some regularization.
Remove/replace features with a high number of NAs.
Remove highly correlating features (keep the best ones), and find features with high multicollinearity.
The last 3 weeks I have played around with the little Python program I made. I find it very helpful to remove features—
For fun, I started with 300 features; the top 10% bucket gave 60% with extremely high turnover. When I scaled it down to the "best" 20 features, the top 10% bucket gave 55% with a very low turnover. So with that little research project, I have a very good base of features to build from with low turnover.