Coresets for Regressions with Panel Data

This paper introduces the problem of coresets for regression problems to panel data settings. We first define several variants with and then present efficient algorithms construct size that depend polynomially on 1/$\varepsilon$ (where $\varepsilon$ is error parameter) number parameters - independent individuals in or time units each individual observed for. Our approach based Feldman-Langberg framework which a key step upper bound "total sensitivity" roughly sum maximum influences all individual-time pairs taken over possible choices parameters. Empirically, we assess our synthetic real-world datasets; coreset sizes constructed using are much smaller than full dataset indeed accelerate running computing objective.