Let's say I have a dataset with some number of features (e.g., 10) and a target variable.

I create 10 PCs from the dataset excluding the target variable. Then, I run a few classification algorithms (XGBoost, SVC, RF, etc.) on these new PC features to predict the target variable.

Is there a reason to expect that this technique will outperform simply using the original features, in the case where we use all available PCs?

I have some intuition that this may help some models learn by making multicollinearity less of an issue, but I'm not sure this is right.

  • $\begingroup$ I'd recommend that you don't use abbreviations in the title. PC is not really a standard abbreviation, so people will not immediately understand the idea of your problem by reading the title. $\endgroup$
    – nbro
    Dec 14, 2023 at 11:35


You must log in to answer this question.