Dimensionality Reduction Algorithm Research For Multi-Label Classification

In multi-label classification, one instance is probably associated with multiple labels, and different labels sometimes are overlapped. Simultaneously, high dimensions of instance probably lead to deteriorate performance of multi-label classifiers, and increase the complexity of computation. Therefore, the research of dimensionality reduction becomes very important.For the importance of reducing dimensionality, two dimensionality reduction methods in multi-label classification are proposed in this thesis:(1) a dimensionality reduction method by minimizing correlations among features and maximizing dependencies between features and labels in multi-label classification (MCMD);(2) an embedding multi-label dimensionality reduction method based on linear ranking SVM (ELRS). MCMD combines features correlations minimization function of principal component analysis and dependencies between features and labels maximization evaluated by Hilbert-Schmidt Independence Criterion to extract a low-dimensional space. Meanwhile, a closed-from expression and balance factor evaluation algorithm of MCMD model are both obtained. In ELRS, we embed feature extraction process in a linear ranking SVM in order to consider labels ranking information and design a fast iterative algorithm for it.In experiments, our paper collects ten multi-label classification evaluation metrics, eight multi-label data sets, five multi-label dimensionality reduction methods and two multi-label classifiers.For MCMD, we demonstrate whether balance factor evaluation algorithm is valid. Afterwards, there is a comparison between MCMD and five dimensionality reduction methods with eight data sets using two multi-label classifiers. Experiments show that MCMD can achieve lower dimensions without deteriorating classification performance.For ELRS, we evaluate the dimensionality reduction rate firstly, and then compare ELRS and five algorithms with eight data sets using two multi-label classifiers. According to experimental results, ELRS also works well.