Detecting abrupt changes in data streams is a classic problem, but it becomes harder when the model linking observations to covariates is nonlinear and the post-change behavior is unknown. The new arXiv preprint considers this setting for generalized linear models, where the goal is to raise an alarm as quickly as possible after a change while controlling false alarms.

The authors propose a variational estimation approach to handle the unknown post-change parameter. Rather than assuming a known shift, the method approximates the needed likelihood quantities, making the detector tractable for GLMs. The abstract notes that such problems arise when monitoring whether the relationship between a learned representation and a downstream outcome has shifted.

Because only one source was provided, there is no independent comparison to report. The contribution appears to be algorithmic: extending quickest change-point detection to a broader model class with unknown post-change parameters, using variational inference to keep computation feasible.