ИСТИНА

Fluorescence excitation-emission matrices (EEMs) combined with parallel factor analysis (PARAFAC) are a powerful tool for studying fluorescence properties of a mixture of fluorophores. One of the most common fields of the use of EEMs decomposition by PARAFAC is the study of distribution, dynamics and transformation of natural dissolved organic matter (DOM). While detailed characterisation of DOM is usually achievable by means of high-resolution mass-spectroscopy, practical constraints (such as inability to ship a mass spectrometer on a polar research vessel or impracticality of taking 5-10 l of seawater samples for pre-concentration) result in the widespread use of less informative, but much more performant optical techniques combined with chemometrics. In this work, 80 samples of DOM from shelf seas were collected during the cruises to the Kara, Laptev, White, and East Siberian seas in autumn (2015-2017) and spring (2016). A few samples were taken from freshwater ponds of Novaya Zemlya archipelago. Since the real world fluorescence data is subject to inner filter effect, scaling problems, and interference from scattering signal, these effects were accounted for by means of, respectively, absorbance-based correction, normalization to Raman peak belonging to water, and interpolation of regions affected by scattering signal. Being an ill-posed unsupervised learning problem, canonical tensor decomposition is hard to perform properly: comparing two solutions is complicated by the fact that the values are defined up to a constant multiplier and can be arbitrarily shuffled. Stability of the PARAFAC solutions is investigated by means of the split-half method and by comparing results of multiple random restarts of the model. Effect of various preprocessing steps and algorithm parameters (such as number of degrees of freedom and termination criterion) on the stability and meaningfulness of decompositions is investigated. Results from different software packages implementing the PARAFAC algorithm (“nwaytoolbox” MATLAB package and “multiway” R package) are compared. Potential to employ line search in addition to already implemented alternating least squares to speed up the decomposition computation in the “multiway” R package is also investigated. Authors acknowledge the financial support provided by RSF project 18-77-00053.