S1 Text -

Section A: hyperparameter values used for different classification models in this work; Section B: The list of most discriminative time series features as obtained by the feature importance analysis on spikebench datasets. Table A: Geometric mean score obtained for the XceptionTime architecture trained on the retinal stimulus classification dataset with alterations in training hyperparameters. Fig A: Metric value evolution during training of an XceptionTime model on the retina dataset with different data preprocessing strategies: blue—no preprocessing, original ISI sequences are used as input; red—standard scaling is performed before feeding the time series to the CNN; green—log-transform (f(x) = log(x + 1)) and standard scaling is applied to the input time series. Top left—training set loss evolution, top right—testing set loss evolution, bottom left—Cohen’s kappa score evolution on the test set, bottom right—test set AUC-ROC evolution during training. One can observe diverging test set loss in cases of no preprocessing or just standard scaling, at the same time training metrics are well-behaved when the log transform is applied to the data. Fig B: Boxplots of tsfresh-extracted feature distributions for features with high discriminative power as detected by the trained decision tree ensemble classifiers in the retinal stimulus type prediction task. A two-sided Mann-Whitney-Wilcoxon test with Bonferroni correction is performed to assess statistical significance; **** denotes p < 1e-4.

(PDF)