With the rapid development of crowdsourcing learning, inferring (integrating) truth labels from multiple noisy label sets, it is also called label integration, has been a hot research topic. And many methods have been proposed for label integration. However, due to the variable uncertainty of crowdsourced labelers, inferring truth labels from multiple noisy label sets still faces great challenges. In this paper we transform the label integration problem into an optimization problem, and exploit a differential evolution-based weighted majority voting method, simply DEWMV, for label integration. DEWMV searches and weights the voting quality of each label through the designed differential evolution (DE) algorithm. In DEWMV, we define three fitness functions, including the uncertainty of the integration label, the uncertainty of the class member probability and the hybrid uncertainty, to search the optimal voting quality for each label. By theoretically analyzing their effectiveness, we choose the hybrid uncertainty as the final fitness function for DEWMV. The experimental results on 14 real-world datasets show that DEWMV is superior to standard majority voting (MV) and all the other state-of-the-art label integration methods used to compare.
CITATION STYLE
Zhang, H., Jiang, L., & Xu, W. (2018). Differential evolution-based weighted majority voting for crowdsourcing. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11013 LNAI, pp. 228–236). Springer Verlag. https://doi.org/10.1007/978-3-319-97310-4_26
Mendeley helps you to discover research relevant for your work.