I'm looking at some time-series data (so unfortunately it cannot be shuffled) and finding the test set has a much wider range than the training set. This means min-max scaling doesn't seem like the best approach. How should I scale the data before modeling to address this kind of issue?
The range of the data (max - min) are not controlling. The purpose of training data is experimental to determine the structure of your propose model; it works as a prototype. Once the prototype of the model is obtain, now you test it. The data used in the training and testing may not be the same, but the differences in range of data (max - min) does not vitiate the proposed model structure obtained from the training because differences in ranges would only change the co-efficient of the variables. The structure of the equation would still remain the same. Both the training and testing model (same model structure) would behave alike; hence, "generalizability."
i also faced with this situation. After i did some inquiries with the other researchers, most of them said that the range of training data need to be the same as test data. Otherwise it considered as extrapolation and the model will not be valid.
I hope I can learn from the answers here on how to approach this problem.
I would recommend not rescaling at all, otherwise, proceed with IRT.
Hope this helps,
Matt
Thank you all!
I think I tend to agree with Paul that the test data shouldn't be too far off to cause serious problems (even if it falls outside). I would definitely agree with Matthew if I could avoid rescaling in the first place. :-)
I have some concerns about Mohd's response though. If we use any information about the test data, wouldn't that be considered data leakage? Can you explain more?
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling