A classifier chain algorithm based on nonlinear feature optimization and automatic generation of label sequence

In the multi-label classification (MLC), each instance is associated with a set of labels. To improve the classification performance, it is important to mine the correlation between labels. The Classifier Chain (CC) algorithm is a well-known algorithm in MLC, which randomly arranges labels in a linear manner. The disadvantage of CC is that the randomization of label sequences results in an unstable classification performance. There is a lack of feature optimization methods for the linear sorting of labels in CC. To solve these problems, a classifier chain algorithm based on nonlinear feature optimization and automatic generation of label sequence (CC-NA) is proposed in this paper. The proposed algorithm provides a more reasonable label sequence by mining asymmetric label correlation. The feature space is optimized by quantifying the features non-linearly and considering the set of labels that CC has learned. In comparison experiments, we chose six benchmark datasets, five of which are in the media domain, where there are frequent correlations between labels. Experiments show that CC-NA has a better classification effect compared with six other advanced MLC algorithms.