Label Distribution Enhancement-Based Label Completion for High-Noise-Ratio Crowdsourcing.

Abstract

Label completion serves as a preprocessing step to handle the sparse crowdsourced label matrix problem, significantly boosting the effectiveness of the downstream label integration. However, existing label completion algorithms generally focus on the crowdsourcing scenarios with low noise ratios but rarely focus on the crowdsourcing scenarios with high noise ratios. In this paper, we focus on the crowdsourcing scenarios with high noise ratios and propose a novel label completion algorithm called label distribution enhancement-based label completion (LDELC). Specifically, we first develop a worker quality estimation method to identify high-quality workers and low-quality workers. Then, for high-quality workers, we develop a class probability estimation method to initialize their label distributions. For low-quality workers, we use the original multiple noisy label sets to initialize their label distributions. Subsequently, we develop a label distribution enhancement method to enhance the low-quality workers' label distributions. Finally, we propagate all label distributions until convergence and then complete the missing labels based on the converged label distributions. Experimental results on both real-world and simulated crowdsourced datasets demonstrate the effectiveness of LDELC.