A(n) Assumption in machine learning

Abstract

The commonly used statistical tools in machine learning are two-sample tests for verifying hypotheses on homogeneity, for example, for estimation of corpushomogeneity, testing text authorship and so on. Often, they are effective only for sufficiently large sample (n> 100) and have limited application in situations where the size of samples is small (n < 30). To solve the problem for small samples, methods of reproducing samples are often used: jackknife and bootstrap. We propose and investigate a family of homogeneity measures based on A(n) assumption that are effective both for small and large samples.