3 ms·
PCA can solve a surprisingly large number of problems. And when it can't directly solve a problem, PCA can be used to as a preliminary analytics step to rank f
by mywittyname 6y ago
PCA can solve a surprisingly large number of problems. And when it can't directly solve a problem, PCA can be used to as a preliminary analytics step to rank features by weight/impact which can then be combined with iterative methods to solve other types of problems.
PCA was my introduction to data science and it remains one of my favorite tools to pull out. I'm really surprised by the number of data scientists that I meet who never use it, or have even never heard of it. Then again, even Data Science From Scratch dedicates like two pages to the subject, so maybe it is not something modern data science programs spend any time on.
- uoaei 6y agoI mostly use it to compress my input features, i.e., to minimize the number of inputs needed on a given model, accepting a tiny loss in accuracy. This is fine because with linear models (such as the first/last layer of an NN) it merely amounts to a rotation, and the mapping back to the original feature space is well-defined.