eCrowd:an embedding representation based on crowdsourcing model
Crowdsourcing is one of the most used methods for data label acquisition.Most existing crowdsourcing algorithms take the crowdsourced results as input and produce the estimate true labels directly,a major drawback of these approaches is that they can't make use of the side information of predictions.In this work,we propose eCrowd,an embedding representation based crowdsourcing model.eCrowd adopts a two-stage working strategy:In the first stage,it first learns the embedding representations of both workers and tasks based on the collected labels,then predicts some specific missing"worker-task"labels with the learned rep-resentations.In the second stage,it produces estimates of the true labels of all tasks,based on the collected and predicted labels.For evaluations,we conduct various crowdsourcing prediction experiments on four real datasets with eCrowd and three other com-parison algorithms,all results shows the superior of our proposed algorithm.
万方数据
