Empirical Bayes large-scale multiple testing for high-dimensional sparse binary sequences

This paper investigates the multiple testing problem for high-dimensional sparse binary sequences motivated by the crowdsourcing problem in machine learning. We adopt an empirical Bayes approach to estimate possibly sparse sequences with Bernoulli noises. We found a surprising result that the hard thresholding rule deduced from the spike-and-slab posterior is not optimal, even using a uniform prior. Two approaches are then proposed to calibrate the posterior for achieving the optimal signal detection boundary, and two multiple testing procedures are constructed based on these calibrated posteriors. Sharp frequentist theoretical results for these procedures are obtained, showing both can effectively control the false discovery rate uniformly for signals under a sparsity assumption. Numerical experiments are conducted to validate our theory in finite samples.