Learning from crowds in multi-dimensional classification domains

Learning from crowds is a recently fashioned supervised classification framework where the true/real labels of the training instances are not available. However, each instance is provided with a set of noisy class labels, each indicating the class-membership of the instance according to the subjective opinion of an annotator. The additional challenges involved in the extension of this framework to the multi-label domain are explored in this paper. A solution to this problem combining a Structural EM strategy and the multi-dimensional Bayesian network models as classifiers is presented. Using real multi-label datasets adapted to the crowd framework, the designed experiments try to shed some lights on the limits of learning to classify from the multiple and imprecise information of supervision.