Back to Volume
Paper: Leave-One-Out Cross Validation: Assessing Consistency of Astrometric Measurements Using “Bayesian Residuals”
Volume: 528, New Horizons in Galactic Center Astronomy and Beyond
Page: 399
Authors: Kosmo O'Neil, K.; Martinez, G. D.; Do, T.; Ghez, A. M.; Chu, D. S.; Sakai, S.
Abstract: Leave-One-Out (LOO) cross-validation assesses how well a data point di within a sample {dj} = {d1, ...,dN} agrees with a Bayesian model fitted to all other data in that sample, {dj}j≠i. This information criteria is mathematically equivalent, at least asymptotically, to eliminating a point and fitting the data N times, but is much more efficient. This diagnostic is useful in orbit or polynomial fitting because it can identify outliers in astrometric measurements that drive the resulting posteriors to improbable regions of parameter space. In particular, several stars in the Galactic center have measured tangential accelerations above a critical threshold such that they are not considered physical. In two such cases, LOO analysis indicates that inconsistent astrometric measurements are, at least in part, responsible for these non-physical accelerations.
Back to Volume