| Author | Message | |||
     Klaas Faber (faber) Senior Member Username: faber Post Number: 49 Registered: 9-2003 |
Hi, I forgot that Sven Serneel's paper deals with multiway data, i.e. the PLS in PLS-DA is so-called trilinear PLS. You will find that work under a similar page that is called "Reliability of multiway calibration". Klaas | |||
| Klaas Faber (faber) Senior Member Username: faber Post Number: 48 Registered: 9-2003 |
Dongsheng, Yes, I believe it is that data set. For interpretation of prediction results, you might consider object-specific prediction uncertainties. [Recall: what CAMO's U-deviation was originally developed for.] These have been calculated for PLS-DA in: S. Serneels, M. Moens, P.J. Van Espen and F. Blockhuys, Identification of micro-organisms by dint of the electronic nose and trilinear partial least squares regression, Analytica Chimica Acta 516 (2004) 1�5 M.L. Griffiths, R.P. Barbagallo and J.T. Keer, Multiple and simultaneous fluorophore detection using fluorescence spectrometry and partial least-squares regression with sample-specific confidence intervals, Analytical Chemistry, 78 (2006) 513-523. Given these uncertainties, it should be straightforward to calculate the probabilities of belonging to a certain class. By the way, this material can be accessed through the following page on my website: http://www.chemometry.com/Index/Links%20and%20downloads/Papers/SelectedReferencesMVC.html Regards, Klaas | |||
| venkatarman (venkynir) Senior Member Username: venkynir Post Number: 62 Registered: 3-2004 |
Dongsheng Bu; Someone may provide clarification between supervised and non-supervised classification. Answer : Imaging you have a big family system like Indian Join family system. Analysis them(here member of the family) based on their ability/contribution to the family is called non-supervised classificaiton. Now move to supversied system, your wife wants you to be dress smart ! like your neighbour. Here the neigbhour's information used as supervisor to you. So when external information used to group ,identifcaiton,classification soon it is called supervsied . When internal parameters of a family used to discrimiante is called no-supervsied. Experts say that Data member is not present and soon . | |||
| Dongsheng Bu (dbu) Advanced Member Username: dbu Post Number: 24 Registered: 6-2006 |
Hi, I have FT-IR measurements of 3 types of vegetable oils as attached. Not sure if Klaas referred to this dataset or not. Spectra were MSC pretreated and truncated. I did similar test as Pierre by choosing (-1, 1), (0, 1) or (0, 100) representing dummy variable in PLS-DA, and got same model and result after scaling, e.g. Regression coefficients need x 100, and prediction value needs x100 from (0,1) to (0, 100). It is interesting that dummy variable can be assigned the way Pierre suggested. Is it for mean centering = 0? I have been in trouble with class membership assignment with PLS-DA method. E.g. in (0,1) case, how do I interpret y_predicted = -0.3, 0.2, 0.45, 0.8, 1, and 1.2, respectively? Probabilities of belonging this class are 0, 0.2, 0.45, 0.8, 1 and 1 respectively, or we need to define a statistic distribution from y_validated in training? Someone may provide clarification between supervised and non-supervised classification. We know PCA models can be used in SIMCA approach (one model for one class). Is it still supervised method because we have training dataset? Also, PCA results can be used in clustering analysis which is a nonsupervised method as I knew. Best regards, Dongsheng
| |||
| Klaas Faber (faber) Senior Member Username: faber Post Number: 47 Registered: 9-2003 |
Eric, It's an example data set for an Unscrambler course so you would need to ask someone from CAMO. Regards, Klaas | |||
| Eric LALOUM (elaloum) New member Username: elaloum Post Number: 3 Registered: 5-2006 |
Dear All, It's very interesting (from a theoretical point of view) that a non-supervised method like PCA could discriminate better than a supervised one like PLS-DA. If there is a linear combination of explanatory variables which can separate perfectly two groups, it's hard for me to figure that a PLS procedure cannot find it. Even if the optimization criterium is not exactly the same between pls and pca estimation, there are not that different. Both are linear methods, modeling in the same hyperplan. On other part, 0 is a very particular value and can be confusing for a dummy variable partly because it's invariant for multiplication. Besides, some times ago, people thought that "0" was even not a true number, but just a useful invention to have coherent theoretical foundation to arithmetic!!! The disjonctive coding 0-1 is of course well founded and used for factorial analysis on qualitative variables but results are not interpreted in the same manner as for a pls regression (and the specific case pls-da). Anyway, I would be very interested to study the data-set reported by Klaas if it's not confidential. Best Eric LALOUM | |||
| Klaas Faber (faber) Senior Member Username: faber Post Number: 46 Registered: 9-2003 |
Pierre, It is difficult to compare PCA and PLS-DA on a conceptual level. While PCA is well-defined in terms of the optimization criterion, PLS-DA is not. The results of PLS-DA even depend on the coding of the "dummy" variables, which doesn't make sense from both a practical as well as a theoretical point of view. [If the choice matters, they are not really "dummy".] Therefore, you can encounter data sets for which you get a perfect separation with 2 PCs while PLS-DA never gets better than 10% misclassification. That's what I once had for a spectroscopic data set. Unfortunately, we seem to be stuck with ill-defined procedures just because that's what is implemented in commercial software. I am pretty much convinced that your statements are correct for DASCO, which is a well-defined procedure. It outperformed, for example, SIMCA in the original publication: I.E. Frank, DASCO - a new classification method Chemometrics and Intelligent Laboratory Systems, 4 (1988) 215-222 I just wonder: has this method been implemented in readily available sofwtare? Best regards, Klaas Faber | |||
| Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 29 Registered: 3-2002 |
Jose, With pls1 and 2 groups we can use any dummy variables. It is common to use 0 or 1. It could be 0 and 100. When the groups have the same number of objects, it is convenient to use -1 et 1. Such a way the limit between the classes is zero. When the number of objects are different, the dummy variables can be n2/(n1+n2) for group 1 and n1/(n1+n2) for the group 2. Then the limit is always zero. The coding for PLS2 is usually 1,0 and 0,1 for 2 groups, 1,0,0; 0,1,0 and 0,0,1 for 3 groups and so on. Winisi codes 2,1 and 1,2 but the results are the same. The limit is than 1.5 instead of 0.5. A test with Unscrambler gives exactly the same Bcoefficients using PLS1 or PLS2 with 2 groups. Concerning the comments from Klaas, PCA is useful to see the data and check if the groups are �homogenous�. If the differences are tiny, PLS2 will give certainly a better discrimination by searching for the �relevant� wavelengths. The directions of main variance used for PCA could mask the differences. If the one or more groups are spread and form subgroups in the PC space, ANN or SVM have to be considered. Pierre | |||
| Jose Miguel Hernadez Hierro (jmhhierro) New member Username: jmhhierro Post Number: 3 Registered: 4-2008 |
Peter, In the equations .Psd of WinISI appear as Estm. min 0, max 1.5 for to 2 classes, the PLS 2 is being developed with 0 and 1 that's not interference posible. Did I get it right? Is it a change of variables in the prediction stage for clasificattion? ie adding a unit Is it here where this interference're talking to? Thanks | |||
| Peter Tillmann (tillmann) Member Username: tillmann Post Number: 12 Registered: 11-2001 |
the reason is simply that ISI (and WinISI) regards "0" as a missing value (standard options). This feature is VERY useful with most multiparameter work, when not all analyses have been done on all smaples, but interferes with DA obviously. I.e. Mark an John did use 1 and 2 to differenciate groups (and 0 for missing). Peter Tillmann | |||
| Klaas Faber (faber) Senior Member Username: faber Post Number: 45 Registered: 9-2003 |
Jose, Good question! The value of the dummy (response) variable can affect the results because PLS is not scale invariant. Default is to use 0 and 1, followed by standardization. When using standardization, the initial values don't matter anymore. By the way, how does a simple PCA scores plot look like? Any useful separation? PLS-DA appears to be one of the standard methods in chemometrics but it is used in a rather "indiscriminate" way itself. I once had a data set where PCA gave perfect separation in 2 dimensions (>99% explained variance) and PLS-DA gave only partial separation with over 5 components. Many will agree that it is always good to start with a PCA. Klaas Faber | |||
| Jose Miguel Hernadez Hierro (jmhhierro) New member Username: jmhhierro Post Number: 2 Registered: 4-2008 |
Thanks Mark I have other cuestion: The PLS method was applied whit the dummy variables, this variables have 0 and 1 values. Did I get it right? Why the class predicton are done in values between 1 and 2? is this a variable change after PLS regression? What is the reason? Why not with variables 1 and 2 directly? Thanks | |||
| Mark Westerhaus (mowisi) New member Username: mowisi Post Number: 3 Registered: 11-2004 |
Yes - - PLS2 with 2 dummy variables is the same as PLS1 with one of the dummy variables. Mark Westerhaus | |||
| Jose Miguel Hernadez Hierro (jmhhierro) New member Username: jmhhierro Post Number: 1 Registered: 4-2008 |
Hi All, I wonder if anybody can help me please. I work with the program WinISI, the algorithm that used to discriminate is always PLS2. If I have two classes PLS2 works like PLS1 because the dummy variables completely correlated.You can get the same information from one variable. Did I get it right? Can I keep this formal statement? Thanks |
