Author |
Message |
David W. Hopkins (dhopkins)
Senior Member Username: dhopkins
Post Number: 193 Registered: 10-2002
| Posted on Tuesday, June 14, 2011 - 12:10 pm: | |
Hi Leonardo, A method that is similar to the cosine or correlation coefficient method mentioned by Howard is the Conformity Index calculation. I have used this with success, and it is relatively easy to implement. See: Plugge, W., C. van der Vlies. The use of near infrared spectroscopy in the quality control laboratory of the pharmaceutical industry. J Pharm Biomed Anal 10 (10-12): 797-803. 1992. Best regards, Dave |
Leonardo Ramirez-Lopez (lramirezlopez)
New member Username: lramirezlopez
Post Number: 4 Registered: 6-2011
| Posted on Tuesday, June 14, 2011 - 8:47 am: | |
Thank you everyone for your replies! Best wishes, Leonardo |
Leonardo Ramirez-Lopez (lramirezlopez)
New member Username: lramirezlopez
Post Number: 3 Registered: 6-2011
| Posted on Thursday, June 09, 2011 - 1:34 pm: | |
Dear Howard Thank you for your reply. The problem with the Mahalanobis distance (MD) in the original spectral space is that the spectral features are highly correlated and this leads to a singular or nearly singular variance�covariance matrix that cannot be inverted and therefore the MD can not be computed. The correlation between samples is a good idea. I think It can be used for identifying samples with similar patterns. I will also try with your cosines idea. Thanks, Leo |
Tony Davies (td)
Moderator Username: td
Post Number: 260 Registered: 1-2001
| Posted on Thursday, June 09, 2011 - 1:31 pm: | |
Hello Leonardo, Welcome to the group! In my work I use a similarity index which is computed as 1/(1-r^2) where r is the correlation between two spectra. If you compare a spectrum with itself you get infinite (1/0) but otherwise you get large numbers when spectra are similar and small numbers when they are very dissimilar. Best wishes, Tony |
Leonardo Ramirez-Lopez (lramirezlopez)
New member Username: lramirezlopez
Post Number: 2 Registered: 6-2011
| Posted on Thursday, June 09, 2011 - 1:18 pm: | |
Dear Jose, Thank you for your reply. You are rigth, I think that it should work for assesing the similarity between groups of samples. However, what I am looking for, is to measure the similarity between instances or samples and not betwwen groups. Some aditional ideas? Thanks again Leo |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 439 Registered: 9-2001
| Posted on Thursday, June 09, 2011 - 12:48 pm: | |
Leonardo - there are several ways to compute a distance between spectra. One way is to compute Mahalanobis Distances without a preliminary PCA, just using the original spectral data. This method has been described in Anal. Chem, 57, p.1449 (1985). Other ways, that use the full spectrum, are computation of direction cosines, and correlation coefficients. \o/ /_\ |
Jose Miguel Hernandez Hierro (jmhhierro)
Intermediate Member Username: jmhhierro
Post Number: 17 Registered: 4-2008
| Posted on Thursday, June 09, 2011 - 11:50 am: | |
Dear Leonardo, For instance you can use the obtanied PCA scores in an hierarchical cluster and then obtain a dendrogram plot Best regards Jose |
Leonardo Ramirez-Lopez (lramirezlopez)
New member Username: lramirezlopez
Post Number: 1 Registered: 6-2011
| Posted on Thursday, June 09, 2011 - 10:01 am: | |
Dear all, I am working in soil NIR spectroscopy. The �standard� method employed for calculating distances between samples is to project the NIR spectra into a principal component (PC) space and then compute the Mahalanobis distance on the PC scores. Do you know some others reliable algorithms for distance or similarity analysis? Thanks in advance, Leonardo Ramirez-Lopez |