It may be caused by Gaussian Process regression. The algorithm tries to fit a curve with the survey data distribution. In the OGS-EGF GPR graph, it shows the dip in the 4d-5d dan ranking.
the old GPR mean rank table
I tried a different set of parameters to relax the boundary limit
https://ntust-mitlab.github.io/go-rank-survey-2020/go-rank-survey-results-2020-gpr-relaxed#A-table-of-rounded-mean-ranks
The curve now is more smooth, but the rank variations and uncertainty of the prediction are also increased. It will be closer to linear regression than gaussian process regression.
the new relaxed-GPR mean rank table
