The quickest change detection problem asks how to spot a change in a data stream as soon as possible after it occurs. In multi-stream settings, an unknown subset of streams may change distribution at the change point, and the optimal generalized likelihood ratio approach is computationally expensive. Existing efficient methods include the rank-weighted sum (RWS) scheme based on Shiryaev-Roberts statistics.
The new paper proposes a Mixture-of-Experts (MoE) framework for this task. Each expert produces a sequence of change-detection statistics, and a change is declared when a weighted sum of these statistics crosses a threshold. The authors apply this framework to multi-stream quickest change detection, yielding a family of algorithms.
Simulation results show that the MoE-based algorithm outperforms the RWS scheme. The experiments cover real-valued Gaussian mean shift data and high-dimensional data generated by Gaussian–Bernoulli Restricted Boltzmann Machines (GB-RBMs). The paper thus offers a more powerful alternative to a leading efficient method, at least in the tested settings.