There are many cases in which I don't know which metric I should use to calculate the value of a feature when the value has a distribution for each data point instead of an absolute number. for example, for the dataset of websites on the web, if the feature is the number of links per page for each website, then we can use multiple aggregation functions such as min, max, median, standard deviation and so to measure it. However, it looks to me that it may result in having many correlated features. So the question is, is it in GENERAL, better to keep multiple aggregations or just choose the best one? Or it most of the times depends on the data. If this is the case, when does it result in duplicated features?
1 Answer
$\begingroup$
$\endgroup$
1
I think if you keep multiple correlated features and then do a PCA to reduce the dimensions, it should choose the ones which are most important for your task. Then you can use the first k principal components for any further work.
