64
J. Eur. Opt. Society-Rapid Publ. 21, 36( 2025)
Figure B1. Preference vs prediction ranks for all images.
Figure
A1. Question asked by the experimenter.
Controlled noise has been added to participant responses by randomly swapping adjacent ranks. Specifically, for each observer and each image, a 20 % probability to swap rank 1 and 2, rank 2 and 3, and so forth was applied. This stochastic perturbation aimed to model intra-observer inconsistencies without drastically altering the underlying preference structure.
To evaluate the impact of these variations, the R 2 value between the prediction and preference ranking by the observers in these cases was computed for 10 independent runs. This variation in the observer ranking led to a R 2 variation of about 0.0031 with a standard deviation of 0.0004, which led to a decrease in model R 2 from 0.973 to about 0.942. This result suggests that the model retains relatively high performance even when faced with moderate inconsistencies in observer data.