Learning from crowds with decision trees

摘要

Crowdsourcing systems provide an efficient way to collect labeled data by employing non-expert crowd workers. In practice, each instance obtains a multiple noisy label set from different workers. Ground truth inference algorithms are designed to infer the unknown true labels of data from multiple noisy label sets. Since there is substantial variation among different workers, evaluating the qualities of workers is crucial for ground truth inference. This paper proposes a novel algorithm called decision tree-based weighted majority voting (DTWMV). DTWMV directly takes the multiple noisy label set of each instance as its feature vector; that is, each worker is a feature of instances. Then sequential decision trees are built to calculate the weight of each feature (worker). Finally weighted majority voting is used to infer the integrated labels of instances. In DTWMV, evaluating the qualities of workers is converted to calculating the weights of features, which provides a new perspective for solving the ground truth inference problem. Then, a novel feature weight measurement based on decision trees is proposed. Our experimental results show that DTWMV can effectively evaluate the qualities of workers and improve the label quality of data.