A Novel Homophily-aware Correction Approach for Crowdsourced Labels Using Information Entropy

Abstract

Crowdsourcing provides a cost effective and conve-nient way for label collection. However, it fails to guarantee the quality of crowdsourced labels. Inspired by homophily in social networks denoting the tendency of individuals with similar char-acteristics to be friends with each other, in this paper we propose a novel Homophily-aware Correction Approach for crowdsourced labels using Information Entropy (namely HaCAIE), to further achieve quality improvement of crowdsourced labels. Specifically, Our HaCAIE can be decomposed into three phases: $i$) seeking full semantic relations among entities, where HaCAIE models multiple explicit and implicit semantic relations among labelers, tasks and categories, based on homogeneous information network and related techniques; ii) calculating homophily, where HaCAIE utilizes adjacent relation matrices of labelers and tasks to calculate homophily among labelers; and iii) correcting labels, where for each task, HaCAIE employs information entropy and constructs a corresponding star homophily network to perform label correction. Our experimental results on six real-world datasets not only show that our HaCAIE performs well, but also demonstrate that HaCAIE can collaborate well with different inference algorithms in the field of crowdsourcing.