Hello, I am a beginner in PARAFAC, and I am following Murphy et al. (2013) and using “drEEM” toolbox to process my data. However, I came up against some questions while dealing with the data and applying the method, and I really hope that I can get some advice from here! Thanks!

  • I am running RANDINITANAL to obtain the least-square model, but it seems like there are two components in my dataset often appear together, and PARAFAC don’t always decompose(?) them into separate components. The output of 100-run RANDINITANAL for 6 and 7-component model shows that there is a chance of 69% and 91% that PARAFAC will treat them as individual components. However, the runs that didn’t decompose them nearly always have smaller SSEs, though the relative difference in SSE is only about 1%, and will be chosen as the least-square model. I’ve read about that “There is no way to say, from the decomposition whether component one is rightfully the first, second or fifth component.” from the online PARAFAC tutorial “Interactive introduction to multi-way analysis in MATLAB (Bro, 1998)”. What about the difficulty(?) for PARAFAC to resolve a particular combination of components during the random process? And is it normal for PARAFAC to resolve a combination of components more easily(?) but with higher SSE? Should I just simply use the output “LSmodel” from RANDINITANAL?
  • Some low-signal samples are included in my dataset, and the contours of the corrected EEMs, especially those with low signal, seem very fragmental. I think this is the reason why I am getting some abruptly-changing excitation and emission spectra. And I think these abruptly-changing spectra are also making my model extremely difficult to validate in split-half analysis. Since I can’t really distinguish the true fluorescence signal from the noise (blank subtraction is done in FDOMCORRECT), removing faulty parts using ZAP or SUBDATASET might not be suitable. I’ve been thinking about smoothing my dataset, however, the instructions of function EEM_SMOOTH in the R package “staRdom” mentioned that smoothing is not advised in PARAFAC analysis. I’m wondering are there any other options when processing these kind of low-signal samples?
  • I’ve read about that the score (concentration) and loadings (spectra) of a component are “only determined up to a scaling (Andersen and Bro, 2003)”, for example, multiplying the excitation spectra by 2 and dividing the emission spectra by 2 at the same time doesn’t change the contribution to the model. What about the relative magnitudes between components? Do the relative magnitudes between components have any mathematical (or physical, or chemical, perhaps?) interpretation? I am asking this because those abruptly-changing spectra mentioned above sometimes feature peaks (or spikes) that have greater value than the spectra of other normal-looking components.
  • If any further explanation for my questions is needed, please let me know. If any of my questions is too basic, or there is any literature I need to read before continuing, please let me know, too.

    Thanks for reading and I really appreciate your time!

    Similar questions and discussions