Abstract

Detecting localized differences between two samples is a central task in scientific data analysis, with applications ranging from signal identification to regime-change detection and model validation. In this talk, I will present EagleEye, a method for identifying local over- and under-densities in multivariate feature spaces. EagleEye detects localized over- and under-densities by comparing the local neighbourhood structure of two samples. Each point is assigned an anomaly score based on whether the composition of its nearby neighbours is consistent with a binomial null model, and these pointwise detections are then consolidated into interpretable anomaly regions. The method also provides estimates of the background level and signal purity of the detected regions. I will first illustrate the method through a synthetic example with known localized over- and under-densities. I will then demonstrate its application in a new-physics search at particle collider experiments in the presence of systematic background mismodelling, and in a climate analysis study of localized changes in spatiotemporal temperature-pattern recurrence. I will also present ongoing work applying EagleEye to searches for faint dwarf galaxy candidates in Gaia DR3 .