About Classifion

Take a tour

Step-by-step example

Workplace infrastructure

Inside the box

Download & Order

Spectrino website

Sicyon website

 Written by and
copyright
Teodor Krastev

 

Presenting following set of images I assume that you have some idea how PCA-MD works .

The workplace

The upper half of the workplace has two pages:

  • Sp.Data (spectral data) is identical to the Spectrino software. It organizes the data, has visualization tools and contains pre-processing options.

  • Interpreter page offers you a monitoring to the commands send to the R-engine and some access to the engine (simple commands could be executed) - that page is similar to what you have under the hood of your car, you don't go there unless there is a trouble.

The lower halt of the workplace contains the modules , in that version of Classifion PCAMD and Classify (later more).

After selecting the base group from Sp.Data hit the "Active group" button to do PCA on it.



Rebuild Spectra page shows you how successful the PCA decomposition was. You select from the list on the right a spec and Classifion will show you the reconstructed (from loadings and scores) spectrum with the residual spectrum. You can follow from here which masses have been more problematic (the biggest peaks in residuals) and look for a reason.



Loadings page is more for completeness, the loading vectors may look like spectra, but they are not and cannot be interpreted as such.



Eigen values page offers you a way to decide how many of the principal components (factors) you should keep. Each eigenvalue is proportional to respective PC significance, so the total eigenvalues threshold defines how mush "significance" you wish to keep in the training. Typically it is from 93% to 99.5% .



Scores page offers a the classical view over the groups (as the clusters here). After you decide how many PC you keep and if they are 1 or 2, that page could help you to exclude some of the specs from the training set (a subset of a group).



Mahalanobis Distance page shows the normalized distance of each spec from the base cluster.



Compactness page is another way to find out the best factor number. It shows how the size of the base cluster varies with factor number calculating the average Mahalanobis distance for each number of factors.



Optimize page will optimize the factor number and will exclude the poor spectra for you. Which preset you will choose depends of what type of error (I or II) you more acceptable for you. The medium correction is the best compromise between the two, soft correction will optimize type I of error more than II and strong correction vice-versa.

Classification

The ultimate purpose of all the adjustments and optimization is to do a real classification. After all the steps sketched above a training file for each group is calculated and saved (manually or by Autopilot) - (.trs). Compiling a list of such a trainings (.trl) is usually more convenient way to classify in batch - feasible here. Now you can classify any selected group of specs. Usually the unknown (substance) is measured more than once, form a group and classify away...

For more information - download Classifion and see the help.

Back to Top