Distributed Partial Inference Under Churn
In this paper, the authors tackle the problem of identifying nodes suitable for participating in inference in the context of a measurement service. They design a simple algorithm, k-Core, to address the problem. They give bounds on the number of measurements with this algorithm. They then develop a distributed version of the algorithm by maintaining and exchanging node and neighbor information at each node. They develop a simulator which implements signaling as well as time step based message passing techniques for the algorithm. They evaluate the performance of both versions of the algorithm with synthetic and experimental churn models. They also vary the inference parameter k.