Table 1. % accuracy of different classifiers applied on different data representations with the best preprocessing combination.

Uncertainty corresponds to 3σ, with σ the standard deviation of the accuracy over the different training sessions. For each classifier, results in bold are significantly the best with respect to a Wilcoxon rank test at a p-value threshold of 0.01 (Lehmann & D’Abrera, 1975 ▸).

Representation	KNN	RF	XGBoost
I(q) space	55.4 ± 0.0	75.0 ± 0.9	73.8 ± 1.1
PCA90	64.9 ± 0.3	69.5 ± 1.6	70.0 ± 1.8
CAE	66.5 ± 1.3	69.6 ± 1.0	70.1 ± 1.5
CNN	84.1 ± 1.1	84.9 ± 1.2	86.7 ± 1.6