Double weighted K-nearest voting for label aggregation in crowdsourcing learning

Abstract

In this article, we propose a new voting strategy in crowdsourcing learning by avoiding the effects of missing labels and poor workers. To achieve this, we apply K-nearest neighbor idea and fitting learning to consider the importance of neighbor sample labels and the importance of workers, respectively. Specifically, we apply different weights to different neighbors based on the distance of the sample neighbors, which is important for the neighbor sample labels. At the same time, we propose an effective worker model to consider the importance of workers by removing redundant workers. In addition, we use an alternate iterative optimization algorithm to solve our proposed model. The experimental results show that the proposed algorithm achieves better performance in label aggregation accuracy than the comparison algorithm.