Abstract
We consider the problem of localizing change points in a generalized linearmodel (GLM), a model that covers many widely studied problems in statisticallearning including linear, logistic, and rectified linear regression. Wepropose a novel and computationally efficient Approximate Message Passing (AMP)algorithm for estimating both the signals and the change point locations, andrigorously characterize its performance in the high-dimensional limit where thenumber of parameters $p$ is proportional to the number of samples $n$. Thischaracterization is in terms of a state evolution recursion, which allows us toprecisely compute performance measures such as the asymptotic Hausdorff errorof our change point estimates, and allows us to tailor the algorithm to takeadvantage of any prior structural information on the signals and change points.Moreover, we show how our AMP iterates can be used to efficiently compute aBayesian posterior distribution over the change point locations in thehigh-dimensional limit. We validate our theory via numerical experiments, anddemonstrate the favorable performance of our estimators on both synthetic andreal data in the settings of linear, logistic, and rectified linear regression.