基于标记感知消歧的偏标记学习算法

Partial label learning algorithm based on label-aware disambiguation

偏标记学习作为一种弱监督学习框架,其目标是从带有噪声标记的偏标记数据中学习一个多分类模型。为解决偏标记学习中标记信息利用不充分且分类效果不佳的问题,提出了一种基于标记感知消歧的偏标记学习算法。通过协同特征空间和标记空间的判别信息来确定示例间的相似程度,并利用示例的相似关系与标记空间中的重构误差来实现消歧过程。在训练分类模型过程中,基于最小二乘损失提出了一个可以同时训练预测模型和消除标记歧义的框架,并采用交替迭代优化的方法获取最佳分类模型。在3组人工合成的UCI数据集和6个真实数据集进行实验,并与现有算法进行对比分析,表明PL-LAD算法具有较好的分类性能表现。

As a weakly supervised learning framework,the goal of partial label learning is to learn a multi classification model from the partial label data with noisy labels.In order to solve the problem of insufficient utilization of label information and poor classification effect in partial label learning,a partial label learning algorithm based on label-aware disambiguation was proposed.The algorithm determines the similarity relationship between instances through the discrimination information of collaborative feature space and label space,and uses the similarity relationship and the reconstruction error in label space to implement the disambiguation process.In the process of training classification model,a framework was proposed based on the least square loss,which can simultaneously train the prediction model and eliminate the labeling ambiguity.And the optimal classification model is obtained by alternating iterative optimization method.Experiments on 3 artificial UCI data sets and 6 real partial label data sets show that PL-LAD algorithm has better classification performance compared with existing algorithms.