| Author | Message | ||
     Barking Gecko |
Dear Sir, I want to clearly discriminate between two spectrally similar products (V1 and V2), and at present am using M Distance within "discriminate" software in Grams32 to attempt this. If I run V1 unknowns against my developed V1 and V2 calibrations, I am finding that the V2 calibrations often come in with a lower M Distance than the V1 calibrations, as they do also for their own (V2) variety. This quite often results in 100% prediction of V2 unknowns but only around 25% prediction of V1 unknowns (75% mismatch to V2). I am not sure whether to try keep trying to reduce the effectiveness of the "interfering" V2 calibrations such that the "correct" V1 calibration will result in a correct match with the lowest M Distance; or attempt to improve the prediction power of the V1 calibration? I have already tried to improve the V1 calibrations, but have had no success with addition of more samples or calibration/pretreatment adjustment. I have also tried reducing the effectiveness of V2 calibrations, but have lost prediction power of V2 also. What do I do next? Please help if you can. | ||
| hlmark |
I love qeustions that say "I need help but can't tell you anything about what I'm doing". At least when you answer them you can't be wrong. But given the lack of information, it's only possible to give very general guidelines. Sounds like with the algorithm, wavelengths and data transforms (if you're using any) you have at least partial overlap between the data represeting the two different materials. GRAMS gives you capability for only some of the many possible approaches. If you're not doing any data transform, you can try some; the best ones to try will dpeend on the nature of the spectra and particularly the differences between them. Derivatives will remove a large part of any physical effects that might be spreading the data out more than necessary, and some other transforms will do the same. You might try different algorithms. Again, which ones are likely to help will depend on the nature of the data from the two materials, and how they relate to each other. Sometimes using fewer wavelengths than the full spectrum will help. The limit of that approach is to use individually selected wavelengths with Mahalanobis Distance computations. Good luck. Howard | ||
| Bruce H. Campbell (Campclan) |
It sounds like you have spectra from only the two types. If that is true, you may be forcing the program to select one, regardless if it is correct or not. Bruce | ||
| Stephen Medlin (Medlin) |
I agree with Bruce in that you may be forcing a classification where one does not exist. Many moons ago, I developed a neural network for Raman interpretation. My results improved when I added a "None of the above" category. Stephen | ||
| hlmark |
I don't think that's the case here, Steve. Our friend Barking KNOWS that his samples represent one of the two materials. But I think I should have clarified a point on my last message. All the suggestions I had were intended to find ways to enhance whatever differences might exist between the spectra of the two materials. If you know somwthing about the spectroscopy involved, that might indicate a direction to take. BTW - it would also be nice if Mr. Gecko would be polite enough to identify h8imself. Howard |