Identify the Best Mechanism-Selection Heuristic

Identify the best heuristic for selecting among the DataSplit, Gaussian, and Thresholdout mechanisms using the adaptivity and query-count estimates produced by AdaptFun, in order to minimize generalization error.

Background

AdaptFun estimates a program’s adaptivity and total number of query requests, and the paper uses these symbolic quantities with asymptotic generalization bounds to choose among DataSplit, Gaussian, and Thresholdout mechanisms. The proposed selection rules compare square-root and logarithmic expressions involving the estimated query count, but the authors characterize these rules as rough heuristics because they omit constants from the underlying theoretical bounds. They explicitly leave open the problem of determining which heuristic most effectively maps AdaptFun’s estimates and the instantiated program parameters to the mechanism yielding the lowest generalization error.