Research On Label Coding Algorithms For Multi-label Classification

In multi-label classification, one instance is probably associated with multiple labels, and different labels sometimes are overlapped. Simultaneously, as the dimensionality of label space grows, it increases the complexity of computation. On the other hand, the sparseness among labels probably leads to deteriorate performance of multi-label classifiers. Therefore, the research of label coding for multi-label classification becomes very important.Label coding approaches mainly are divided into label compression, label expansion and hybrid reduction methods. From view of label compression, two label coding methods for multi-label classification are proposed in this thesis:(1) a label compression coding approach through maximizing dependence between features and labels (LCCMD); (2) a label compression coding approach based on symmetrical network (LCCSN). The former maximizes dependence between features and labels evaluated by Hilbert-Schmidt independent criterion to extract a low-dimensional label space in which fewer learning tasks are conducted. In the latter, we first construct a five-layer symmetrical network, which consists of encoder network and decoder network, and is trained by the idea of extreme learning machine.In experiments, our thesis collects twelve multi-label data sets, ten multi-label classification evaluation metrics and three existing label coding approaches. For LCCMD, we conduct experiments on ten benchmark data sets, and using five different compression rates. For LCCSN, we tune the parameters on training data set before comparison. Experimental results demonstrate that our proposed methods improve performance of classification.