基于多标签随机游走的选择性集成方法用于表情识别

A Selective Ensemble Method Based on Multi-Label Random Walk for Expression Recognition

为了提升分类器集成的性能,本文提出了一种基于多标签随机游走的选择性集成方法,该方法将分类器选择问题建模为多标签分类问题,以灵活有效选择分类器。首先在训练集样本与分类器序列间建立映射,将能正确识别样本的分类器序列视为样本标签,对于一个特定的测试样本,寻找它的近邻样本,并构建出多标签随机游走图,执行随机游走过程,根据收敛后的概率向量选择出对应的分类器序列进行集成。在Fer2013、CK+和JAFFE人脸表情数据集上进行实验,并与一些当前先进的选择性集成算法进行对比,实验结果证明了该算法的可行性和有效性。

In order to improve the ensemble performance of classifier ensemble, a selective ensemble method based on multi-label random walk is proposed, which models the classifier selection problem as a multi-label classification problem to select classifiers flexibly and effectively. First, a mapping is established between the training set samples and the classifier sequence, and the classifier sequence that can correctly identify the sample is regarded as the sample label. For a specific test sample, its neighbor samples are found, and a multi-label random walk graph is constructed. A random walk process is performed, and the corresponding classifier sequence is selected for ensemble according to the converged probability vector. Experiments are carried out on Fer2013, CK+ and JAFFE facial expression datasets, and compared with some current advanced selective ensemble algorithms. The experimental results demonstrate the feasibility and effectiveness of the algorithm.