Information-Directed Random Walk for Rare Event Detection in Hierarchical Processes

TL;DR

This paper introduces an information-directed random walk strategy for efficiently detecting rare anomalies in hierarchical data streams, achieving asymptotic optimality and logarithmic sample complexity under certain conditions.

Contribution

It proposes a novel sequential search policy using an information-guided random walk on tree-structured observations, with proven asymptotic optimality and sample complexity results.

Findings

Asymptotic optimality in detection accuracy.

Logarithmic-order sample complexity in search space size.

Conditions for sublinear scaling in Bernoulli case.

Abstract

The problem of detecting a few anomalous processes among a large number of data streams is considered. At each time, aggregated observations can be taken from a chosen subset of the processes, where the chosen subset conforms to a given tree structure. The random observations are drawn from a general distribution that may depend on the size of the chosen subset and the number of anomalous processes in the subset. We propose a sequential search strategy by devising an information-directed random walk (IRW) on the tree-structured observation hierarchy. Subject to a reliability constraint, the proposed policy is shown to be asymptotically optimal with respect to the detection accuracy. Furthermore, it achieves the optimal logarithmic-order sample complexity with respect to the size of the search space provided that the Kullback-Leibler divergence between aggregated observations in the presence and the absence of anomalous processes are bounded away from zero at all levels of the tree structure as the size of the search space approaches infinity. Sufficient conditions on the decaying rate of the aggregated observations to pure noise under which a sublinear scaling in the size of the search space is preserved are also identified for the Bernoulli case.