Feature Selection Accuracy Comparison

[ericborgos]

It would be helpful if on the page where you give the Titanic examples, if you also displayed a comparison of the accuracy scores, to show which feature selection methods were the most effective.

[msat59]

@ericborgos , no single method is the best in all cases; it depends on the specific problem. Additionally, the dataset size plays a significant role; you can't judge by just using Titanic dataset. I tested various methods on a small synthetic dataset (using sklearn.make_regression), and GrootCV successfully identified all informative features. However, it performed poorly on a real-world problem with over 5000 features. In that case, Leshy, using both native and SHAP importance, produced much better results than other methods. I later did hyper-param tuning, calculated SHAP values and kept top 50% features. I did it in several iteration until I had around 50 features (desired by the business).