基于矩阵填充的众包学习模型研究

本文提出一种鲁棒低秩近似算法(ROLA)来学习标注者之间潜在的相似性,进而解决标注数据集中的噪声。ROLA通过构造一个低秩矩阵模型,来捕获标签中的潜在相关信息,与问题的潜在特征向量。实验结果表明,ROLA在四个数据集上的准确率最高。并且与现有算法相比,在优化时间上也存在相应优势。

Research on Crowdsourcing Learning Model Based on Matrix Filling

This paper proposes a robust low rank approximation algorithm (ROLA) to learn the potential similarity between annotators and to solve the noise in annotated data sets. ROLA constructs a low rank matrix model to capture latent correlation information in tags and latent eigenvectors of problems. The experimental results show that ROLA has the highest accuracy on four data sets. Compared with existing algorithms, it also has corresponding advantages in optimization time.