I am doing a SIMCA classification modelling using the Unscrambler software. The usual approach is to run a global PCA first to ascertain existing classes are clearly distinguishable before continuing. What do i do after running the global PCA and dicover my classes are not easily distinguishable but having overlaps. Do i continue with SIMCA modelling all the same or what step should i take from there? I am somehow hooked!
If results of PCA global model do not provide distinguishable classe it is necessary to revise your data set in order to identify redundant variables.
If this revision is not possible by using visual inspection I recommend the use of methods for variable selection.
Even if these strategies do not work I suggest the use of another method of classification.
PCA just performs a linear transformation to reduce the dimensionality of your data into it's principal components, so if the data was not linearly separable to begin with, it will not be afterwards. I don't know what your data looks like, but I would give it a try to some non-linear method like ISOMAP. There is a ton of these methods, take a look https://en.wikipedia.org/wiki/Nonlinear_dimensionality_reduction
Other method you could try would be some type of neural network.
Dear Adeyemi,
I am not sure if you could provide details of your samples. If I knew more about your samples I would be able to give you some step by step examination of your data. My experience is with FT-NIR spectral data, are you using NIR spectroscopy? If you want to reach me send me an email at nirtechnologies.com
All the best,
Hormoz
In a real SIMCA analysis you model your classes separately by separate PCA models. An overall PCA model gives you a nice idea of how easy (or not) your classification problem will be but it doesn't give you the final answer about that, because it's a different method. From what you write it sounds like there is some kind of separation between classes but not perfect, since they are overlapping. Why don't you just try your SIMCA model?
Thank you Marta. I did try it but no good membership grouping. Most practice that i have seen give good membership grouping are those that the classes were easily distinguishable right from the global PCA stage. This means indeed that my classes has a great overlap and was thinking i need to deal with this as much as possible before progressing. Thank you once again.
Dear Adeyemi,
I totally agree with Marta's comments. I have one further question and a suggestion:
- Which index did you use to define the class membership?
- Why do not you try Partial Least Squares Discriminant Analysis (PLSDA), which will focus directly on the variation mostly related to the class differences?
Best regards,
Raffaele
Adeyemi Adegbenjo
Could you maybe give me some instructions about which literature did you use in order to figure out how exaclty to perform SIMCA method? Also, which software do you use? Do you think Unscrambler is okay?
Thank you very much
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification