基于机器学习辅助的人机数据融合方法

Human-machine Data Fusion Method Based on Machine Learning Assistance

从未经验证的数据中获得较高的质量的回答是一个经典问题。考虑了一种存在历史数据的二分类任务,利用历史任务训练出分类器后通过比对工人以及分类器给当前任务提供的回答估测分类器的准确率以融合数据来提高最终结果的准确度并且节约了劳动力。实验得出,在已知正负标签的先验概率时,充分利用先验概率进行的数据融合可以将分类器与工人的回答很好的融合并且获得更加准确的结果。实现了在没有正确的标签的情况下,训练分类器会根据不同的情况融合分类器与工人回答来提高最终结果的准确度的目标。

Information Elicitation without Verification(IEWV) is a classic problem. The article considers a binary classification problem with historical tasks. After the classifier is trained by the historical tasks, the accuracy of the classifier is estimated by comparing the answers provided by the worker and the classifier of the current tasks, and the final results are obtained by merging the data to improve the accuracy of the final results and save labor. Obtained from the experiment, making full use of the known prior probability of positive and negative label can fuse the classifier and workers’ answers well and obtain more accurate results. It achieves the goal of combing the answers provided by the classifier and workers in different situations and improving the accuracy of the results without the correct labels.