I have a training set of time series data where each feature is highly skewed with long right tails. I get poor classification performance using regression and neural networks. Only tree-based methods give good classification performance, presumably because they're not affected by the skewness. I have tried using transformations like log/sqrt/boxcox but with little success. After transformation the features still look very skewed.
I was wondering whether there is a way to use machine learning to automatically learn the best transformation to apply to the feature, or maybe learn how to truncate a feature whilst still retaining the important information?