Label Aggregation with Instance Grouping Model

Abstract

Label aggregation is one of the key topics in crowdsourcing research. Most researchers make their efforts in modeling ability of users and difficulty of instances. In this paper, we consider label aggregation from the view of grouping instances. We assume instances are sampled from latent groups and they share the same true label with their corresponding groups. We construct a graphical model named InGroup(Instance Grouping model) to infer latent group assignment as well as true labels. The experimental results show the advantages of our model compared with baselines.