Attribute Reduction For Multi-label Classification Based On Positive Region

On the basis of the rough set theory,this paper proposes four attribute reduction methods for multi-label classification problem and designs corresponding algorithms.At the same time,in order to improve the efficiency and speed of calculation,we evaluate the importance of attributes by using the lower approximation of the label information set instead of the decision class.Then the relationship between the proposed methods and the two classical attribute reduction methods is analyzed.The experimental results show that the proposed methods are more suitable for multi-label classification problems.These four attribute reduction methods are mainly divided into two types.In the third chapter,we first give the definition of the label information set.The multi-label classification problems can be converted to a series of binary classification problems by the label information sets.Then the label positive region reduction method and neighborhood label positive region reduction method were put forward,and their properties were discussed.Finally,the corresponding algorithms were designed.Some comparison experiments were carried out on the actual multi-label data sets.The experimental results illustrated the validity and feasibility of the two attribute reduction methods.If a sample belongs to the lower approximations of multiple label information sets at the same time,where the cardinality of the positive region were calculated by the label positive region reduction method,the cardinality of the positive region only increases 1.The multi-label information of the sample is lost,therefore,the cardinality does not fully reflect the classification ability of the attribute subset.Chapter 4 presents label dependency reduction method and neighborhood label dependency reduction method.The experiment results show that the two mentioned methods can remove redundant attributes without reducing classification accuracy for most multi-label data sets.