| Author | Message | ||
     venkatarman (venkynir) Senior Member Username: venkynir Post Number: 93 Registered: 3-2004 |
I have written NIPALS routine in object oriented langauge based on the publication of Paul Geladi and Bruce R.Kowalski Analytica Chimica Actam 185(1986) -1-17. Since our variable y is one component , we bypassed step 5 to 8 as suggested . Our predicted value for traning set well matchs with true value. This also verfied with PLS tool box and unscrambeler package. Kindly suggest a method to(y) predict unknown spectra. we have stored the value of P,T W & b. Is the mean centre is necessary for pre-processed NIRS data ?. | ||
| Daniel Alomar (dalomar) New member Username: dalomar Post Number: 3 Registered: 2-2009 |
OK. Thanks a lot, Jos� Ram�n. Daniel | ||
| Jos� Ram�n Cuesta (jrcuesta) New member Username: jrcuesta Post Number: 4 Registered: 11-2009 |
Hi, I have found some documentation about this type of validation from ISI: What Step up or stepwise really does is a leverage correction that computes what the prediction residual would be if one sample is omitted from the calibration. The problem is that this sample is not remove from the wavelength selection which is the main cause of overfitting. So in most of the cases is not very usefull. So external validation is better Best regards J. Ram�n | ||
| Daniel Alomar (dalomar) New member Username: dalomar Post Number: 2 Registered: 2-2009 |
Thank you Pierre and Ram�n, It�s weird, because when the software is performing the calculations (stepwise) a window in the upper margin flickers with �performing cross validation� but, if any, it works as a black box... Best wishes Daniel | ||
| Jos� Ram�n Cuesta (jrcuesta) New member Username: jrcuesta Post Number: 3 Registered: 11-2009 |
Dear Daniel and Pierre, I agree with Pierre in the way to develop the calibration for MLR stepwise in Win ISI and is what I do when I develop a calibration with this method,... but I ask exactly the same question than Daniel to ISI some time ago and they reply to me that there are some calculations for these statistics: The SECV and 1-VR are calculated by leverage correction residuals that in the case of MLR is similar to a cross validation "leave one out". Best Regards J. Ram�n | ||
| Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 45 Registered: 3-2002 |
Dear Daniel, Do not worry about that. This is just a small bug. There is no CV with the stepwise. You need a test set with the stepwise method. The most important is that the model and the predictions are calculated correctly. Pierre | ||
| Daniel Alomar (dalomar) New member Username: dalomar Post Number: 1 Registered: 2-2009 |
Dear all, I performed a calibration with Winisi II, global calibration, with selected wavelengths, in stepwise. After it was completed, I selected �graphs� and �statistics�. And, apart from the plot, a box appeared with data from crossvalidation (SECV and 1-VR). But no crossvalidation options are present in the previous menus and none was apparently performed. Am I missing something or it�s a bug in the software? Any hints from Winisi users or developers? |