Big data sketching with model mismatch

Data reduction for large-scale linear regression is one of the most important tasks in this era of data deluge. Exact model information is however not often available for big data analytics. Therefore, we propose a framework for big data sketching (i.e., a data reduction tool) that is robust to possible model mismatch. Such a sketching task is cast as a Boolean min-max optimization problem, and then equivalently reduced to a Boolean minimization program. Capitalizing on the block coordinate descent algorithm, a scalable solver is developed to yield an efficient sampler and a good estimate of the unknown regression coefficient.