一种基于Tri-training的众包标记噪声纠正算法

A Tri-training-Based Label Noise Correction Algorithm for Crowdsourcing

在众包学习中,使用标记集成算法得到的集成标记中仍然存在一定程度的标记噪声.本文受三重训练思想的启发,提出了一种基于tri-training的众包标记噪声纠正算法(Tri-Training-based Label Noise Correction,TTLNC).TTLNC首先使用过滤器获得干净集和噪声集,然后在干净集上进行bagging分别训练三个不同的分类器,并通过这些分类器重新标注噪声集中的实例,同时按照实例分配策略将实例分配给相应的训练集.最后在新训练集上重新训练三个不同的分类器,并用新分类器的分类结果重新标注所有实例.在仿真标准数据和真实众包数据集上的实验结果表明TTLNC比其他四种最先进的噪声纠正算法在噪声比和模型质量两个度量指标上表现更优.

In crowdsourcing learning,a certain level of label noise still exists in integrated labels obtained by employing ground truth inference algorithms.Inspired by the tri-training idea,this paper proposes a tri-training-based label noise correction(TTLNC)algorithm for crowdsourcing.TTLNC at first employs a filter to get a clean set and a noisy set and then trains three different classifiers from the bagged clean set.Furthermore,each instance from the noisy set is relabeled by these classifiers and assigned to the corresponding training set according to the designed instance assignment strategy.Finally,three classifiers are retrained on three new training sets and are used to relabel all instances.Experimental results on both simulated benchmark data and real-world crowdsourced data show that TTLNC significantly outperforms other four state-of-the-art noise correction algorithms in team of the noise ratio and the model quality.