Photo-z in cosmological surveys with ML algorithm: Effects of incompleteness

One of the crucial keys in the cosmological studies is the estimation of an accuracy redshift of a large number of galaxies.
The selection of spectroscopic training sample is one of the most important features to obtain a precise photo-zs. Unfortunately, the science samples may contain deeper magnitudes that are not well represented in the spectroscopic samples. These differences raise doubts about the confidence of the photomectric redshift provided by the algorithms.

DNF is a nearest neighbor algorithm for determing the photometric redshift. We present the effects to train this algorithm with an incomplete sample. We provide a method to detect and measure the incompleteness. We apply this method to the Y3 DES Deep Fields catalogue and compare the results with those obtained from template approaches.