One of the most commonly misused statistical technique is scaling. Most of the time, we normalize the data with the knowledge that we don't process in the first place from training data. I recently found LIBSVM a good tool doing SV machinery. However, I am not so sure if I agree with the scaling data in the model. In the "A practical guide" APPENDIX B, the author states a common mistake in scaling training and testing data
stresses the importance of using the same scaling factors for training and testing sets?
Does every one here agree with this statement?
Best Answer
Support vector machines do their magic through measures of distance. Changing the scaling is equivalent to redefining distance. This is why an SVM model requires the exact same scaling factors for both training and testing data.
I certainly do.