Robust crowdsourcing-based linear regression

In most machine learning problems, the labeling of the training data is an expensive or even impossible task. Crowdsourcing-based learning uses uncertain labels from many non-expert annotators instead of one reference label. Crowdsourcing based linear regression is an efficient way for function estimation when many labels are available for each instance. However, methods in literature have a poor performance against large noise and outliers in labels. To tackle this problem, we proposed a novel robust crowdsourcing-based linear regression derived from information theoretic learning. The proposed problem is not convex, but it can be efficiently solved by half quadratic programming. The proposed model has a close relation with weighted crowdsourcing-based linear regression, in which labels of each annotator weight adaptively and iteratively. The Performance of the proposed method evaluated on several artificial data sets in different circumstances. Experimental Results demonstrate the efficacy and robustness of the proposed method.