I have gotten some advice from a PhD statistician on doing predictive modeling on large datasets (lots of variables AND lots of observations) that I should perform transformations to eliminate skewness in my numeric variables. I should first try various power transformations, then when that doesn't work, try rank transformations, then when that does not work to bin my variable into ordinal categories or in extreme cases (e.g. >95% of a variable is equal to zero) turn the variable into a nominal variable of some sort. This is all with the goal of being able to fit various types of models to the data (Neural Nets, SVMs, Logistic Regression, etc.) I am having trouble finding any advice backing this up. I have found that normalizing, scaling and box cox transformations seem to be common ways to improve models, but not this. Can someone help validate/invalidate this advice?
1 Answer
You didn't give any context, and you don't say if the advice was about transforming response variable or predictor variable. At that level of generality, this cannot be sound advice.
That is not to say that transformations should not be applied, but it is not a silver bullet. To often we see (for instance on this site) people who want to transform in a way that do not make sense, maybe even destroying interpretability of model. If you believe some transformation could be a good idea, try it, and then validate the model and predictions to see if it really helped.