Augmenting photometric redshift estimates using spectroscopic nearest neighbours

As a consequence of galaxy clustering, close galaxies observed on the plane of the sky should be spatially correlated with a probability that is inversely proportional to their angular separation. In principle, this information can be used to improve photometric redshift estimates when spectroscopic redshifts are available for some of the neighbouring objects. Depending on the depth of the survey, however, this angular correlation is reduced by chance projections. In this work, we implement a deep-learning model to distinguish between apparent and real angular neighbours by solving a classification task. We adopted a graph neural network architecture to tie together photometry, spectroscopy, and the spatial information between neighbouring galaxies. We trained and validated the algorithm on the data of the VIPERS galaxy survey, for which photometric redshifts based on spectral energy distribution are also available. The model yields a confidence level for a pair of galaxies to be real angular neighbours, enabling us to disentangle chance superpositions in a probabilistic way. When objects for which no physical companion can be identified are excluded, all photometric redshift quality metrics improve significantly, confirming that their estimates were of lower quality. For our typical test configuration, the algorithm identifies a subset containing similar to 75% high-quality photometric redshifts, for which the dispersion is reduced by as much as 50% (from 0.08 to 0.04), while the fraction of outliers reduces from 3% to 0.8%. Moreover, we show that the spectroscopic redshift of the angular neighbour with the highest detection probability provides an excellent estimate of the redshift of the target galaxy, comparable to or even better than the corresponding template-fitting estimate.