I would like if there are work (publications, extensive results, recommendations) about the SVM feature normalization when the instance features are histograms.
Previous work: In this question https://stackoverflow.com/a/7863715, they talk about the chi square distance (commonly used in histograms) and feature normalization but I don't know if the features are histograms. Moreover, this paper http://www.cs.huji.ac.il/~ofirpele/QC/ mention that chi square is sensitive to feature normalization (don't specify positive or negative performance).
I think that the common normalization between [0, 1] and [-1, 1] can affect the histogram information.
