Research On Multi-label Learning Algorithms Based On Samples And Label Correlations

Multi-label learning is a kind of framework used to solve the problems of ambiguity object modeling. Its research results have been widely applied in many fields, such as text classification,web page classification, scene classification, etc. It has become a hot research topic in machine learning. Many learning algorithms have already emerged in multi-label classification problems,including methods based on k-nearest neighbors, support vector machine(SVM)or decision tree.And these algorithms can be classified into two categories: algorithm adaptation methods(AAM)and problem transformation methods(PTM). As the name implies, algorithm adaptation methods deal with multi-label data at the algorithm level. Similarly, the problem transformation methods start from the multi-label learning problem, and then transform it into several single-label learning problems.Firstly, this paper introduced the background and significance of multi-label learning briefly,and the research status of multi-label learning at home and abroad. Secondly, we introduced multi-label learning theory, including the definition of multi-label learning problems, the introduction of multi-label learning strategy, and the commonly used theory in multi-label learning, such as evaluation criteria, specific multi-label learning algorithms and data sets, etc. In the process of multi-label classification problem, the most important problem is how to take advantage of the trained classifier to divide samples into predefined categories effectively. This paper analyses from two aspects: label correlation and attribute feature selection in multi-label classification. The main research results are as follows:(1) Based on the research on label correlation problems existing in the multi-label learning,it proposed a modified multi-label learning algorithm with neighborhood rough sets. The model of neighborhood rough sets is introduced in the multi-label learning, then a new learning framework named MLRS is constructed by researchers. We can take advantage of neighborhood rough sets to find out all possible related labels, and eliminate all irrelevant labels for a given sample. According to the neighborhood and the relationship between different labels, we can predict the correct label scope for the sample. However, if the positive examples of a class in the boundary is too little, it will induce a little scale occupied by this class in neighborhood. At this time, if we only consider the number, it will be easy to cause wrong points. To address this issue,this paper has made a corresponding improvement based on MLRS. In the border area, we can establish a mapping relationship according to the Euclidean distance and their numbers, which between samples and the different samples in neighborhoods. And then we can take advantage of the mapping relationship to predict the labels of test sample. Experimental results show that the proposed method can improve the classification performance of the algorithm.(2) Based on the research on feature selection problems existing in the multi-label learning,a modified weighted classification algorithm with label specific feature for multi-label learning is proposed. An object has a variety of semantic information, because it also contains attributesabout its semantic. Therefore, the effective transformation of attributes which depicts the relationship between samples and labels, will make the multi-label learning process more reliable.LIFT is a multi-label learning algorithm based on the label specific feature. This paper proposed a new method named LIFT-LOCW aiming at the shortcomings of LIFT, this algorithm has improved the classification accuracy by using the method of weighted. Experimental results show that the proposed algorithm performs better than the other commonly used multi-label learning algorithms.