'Ask our experts' program gives you the opportunity to ask questions to our panel of experts. Our expert panel offers you tips, tricks and creative solutions for solving your multivariate problems. The program is designed to help data analysts to choose the right tools and methods to suit their needs. This is an opportunity for you to get advice from experts on data analysis.

Post a question on topics of your interest, and our experts will address it on this page. Selected questions will be treated in further detail through our upcoming webinars.

When using PLS regression, is it possible to override the suggested number of components? For example, cros validation suggests that I use 3 components. Can I force the algorithm to use either 2 or 4? (Larry Bruce) July 16   Answer In The Unscrambler X (and 9.8) one can use a different number of factors during the prediction step, by changing it on the Prediction dialogue.In The Unscrambler one can scroll through the regression results for different numbers of components using the arrow keys in the toolbar. To change the number of factors in a model in The Unscrambler 9.8, in the Regression window go to File->Properties, and on the Model tab, change the number of components. Then save the model. In addition if you wish to calculate the uncertainty test with a different number of factor you can select the option "Number of factors to use" in the Validation tab. 0 votes Can unscrambler make use of multi-core processors? For example, on a quad core machine, would the processing be distributed over all four cores, or confined to one (and therefore to only 25% of the potential processing power)? (Ryan Anderson) July 7   Answer The Unscrambler X does not take full advantage of multicore processors due to its numerical calculations. But the performance of The Unscrambler X will be better in multicore processor system than with a single core processor due to the nature of processor architecture. 0 votes Is there a way to run discriminant analysis using both spectra and lab values form my samples? It seems you can run PCA on spectral data, and PLS on Lab data. How can I incorporate the two and have the software sort out what is "good" vs. "bad"? (Carol Zrybko) May 12   Answer This can indeed be done in The Unscrambler, where in one created indicator variables (or class identifiers) to develop a training set of "good" vs "bad". First create a category variable "Edit-Insert-Category variable" and define the level manually. Then you can use the function to "split category variable" and make columns of 0 and 1. In this special case you have only two values for the category variable so you may want to just create a new variable and code it yourself -1: bad and 1: good. This unique variable can be the Y variable. It is used in the PLS regression. For the predictors (X table) you can have both the spectra and lab values. However you will need to use a scaling as lab values may not be at the same scale as the spectra. You can simply use a weight of 1/std for all variables, even though this is not recommended when using only spectral data. You can probably make a selection of the important variables in a second step. 0 votes I understand that The Unscrambler® 9.8 can use Absolute Pearson Correlation distance in k-means clustering. How does it recalculate centroids in this case? (Andrew Kossenkov) May 12   Answer The calculation of the centroid is independant from the chosen distance. It always takes the barycenter of the group of sample in the cluster. 0 votes We melt steel scrap to produce various grades of varying acceptable concentrations of residual elements. We have our scrap recipes grouped into various mixes of about 5 different grades of scrap. Scrap is inherently inconsistent, but we try to blend and optimize our costs. We would like to take data from the scales of actual weights used in our batches of different commodities and compare with final melt-in chemistry and calculate approximate residual levels, by commodity and use these to adjust scrap mixes, real-time. What product, if any would be best? (Mark Trapp) April 8   Answer First you can run some PLS to model the relationship between the variables of the process and your cost and quality.When you get results from those data you may be able to find some variables that have more influence than the rest. From there you can run a series of experiments based on DoE to be able to optimize your process or find an area of minor variability in your responses.Both analysis can be done with The Unscrambler. 0 votes 1 | 2 | 3 | 4 | 5 |

Panel of Experts

	Brad Swarbrick PAT & QbD		Dr. Frank Westad Methods & Algorithms		Dr. Marion Cuny Sensometric & DoE
	Brad Swarbrick is a PAT specialist consultant with CAMO Software and has over 10 years experience in the application of chemometrics techniques to Mid and Near Infrared Spectroscopic analysers...		Dr. Frank Westad 's working experience over the years includes positions as Research Scientist with SINTEF, Consultant with IDT GmbH (Germany), CSO with CAMO and Senior Research Scientist ...		Dr. Marion Cuny, Ph.D. is a food science engineer and specializes in signal interpretation in NMR spectroscopy applied to food authenticity issues. Awarded a Ph.D in analytical chemistry (2008) from AgroPArisTech ...
	Dr. Katherine A. Bakeev PAT & Near infrared		Suresh Kumar Spectroscopy
	Katherine’s expertise is in the development and implementation of spectroscopic solutions for on-line measurement in manufacturing (PAT, near-infrared and UV-vis spectroscopy) for various ...		Mr. Suresh Kumar, has an M.Sc. in Physical Chemistry and is currently, pursuing his Ph.D. Suresh has more than 8 years of experience in the Quality Control, Analytical Research & Development of food, analytical ...