Abstract

With the emergence of search engines and crowdsourcing websites, machine learning practitioners are faced with datasets that are labeled by a large heterogeneous set of teachers. These datasets test the limits of our existing learning theory, which largely assumes that data is sampled i.i.d. from a fixed distribution. In many cases, the number of teachers actually scales with the number of examples, with each teacher providing just a handful of labels, precluding any statistically reliable assessment of an individual teacher’s quality. In this paper, we study the problem of pruning low-quality teachers in a crowd, in order to improve the label quality of our training set. Despite the hurdles mentioned above, we show that this is in fact achievable with a simple and efficient algorithm, which does not require that each example be repeatedly labeled by multiple teachers. We provide a theoretical analysis of our algorithm and back our findings with empirical evidence.