The binomial distribution models independent success-or-failure observations with a shared probability. In validation workflows, binomial(n, 1/n) lets each observation be sampled independently during bootstrap instead of drawing a full multinomial sample. section 7.4
For datasets much larger than 100, binomial(n, 1/n) closely approximates Poisson(1). Because Poisson(1) does not depend on dataset size, the sampling can be parallelized, reportedly accelerating computation by roughly 100–1,000 times. section 7.4
Bootstrap recall bounds
Independent bootstrap samples produce a recall distribution from which a 99% lower confidence bound can be selected; the reported example places recall above 0.071 on average in 99 of 100 cases. section 7.4
Distribution recall with every point being a bootstrapped original dataset; the red line is a 99% lower confidence bound.
Use the recall-distribution figure to connect repeated sampling with confidence-bound selection. section 7.4
