Semi-supervised clustering with inaccurate pairwise annotations

Pairwise relational information is a useful way of providing partial supervision in domains where class labels are difficult to acquire. This work presents clustering model that incorporates pairwise annotations the form must-link and cannot-link relations considers possible annotation inaccuracies (i.e., common setting when experts provide supervision). We propose generative assumes Gaussian-distributed data samples along with generated by stochastic block models. adopt maximum-likelihood approach demonstrate that, even weak inaccurate, accounting for significantly improves performance. observe also helps detect meaningful groups real-world datasets do not fit original data-distribution assumptions. Additionally, we extend integrate prior knowledge experts’ accuracy discuss circumstances which use this beneficial.