Research On Multi-Label Dimensionality Reduction Based On Label Related Features

In recent years,with more and more multi-label learning problems appear in the practical application scenario,research on multi-label problem has gradually become central issue in those related fields.With the prominence of problems in the multi-label domain such as high dimension,magnanimous redundant information,etc.,the position of dimensionality reduction research in these related domains is becoming more and more important.Although existing multi-label dimension reduction methods have already obtained excellent effect,but most of them do not take into account those differences existed between different labels.Just like relevant literature points out that each different class label may have its own unique attributes and characteristics(i.e.label-specific feature),which are only related to corresponding label,and will be more helpful to identify corresponding label.Based on above theory,in this thesis research work is carried out on the target how to extract features in label-specific feature spaces,then two new algorithms are proposed for implementing dimensionality reduction in multi-label domains.In this thesis,our main research works are as follows:(1)Based on the idea of random subspace method,a new algorithm called LIFT_RSM is proposed to implement feature extraction in label-specific feature space under multi-label conditions.This proposed algorithm refers to design idea of LIFT algorithm,then uses random subspace methods to mining potential association information in each generated label-specific feature space,and combines those information contained in pair-wise constraints used to extract effective low dimensional features.At last,results of contrast experiments show that proposed algorithm LIFT_RSM could effectively improve and optimize those related effectiveness of multi-label learning.(2)A new multi-label algorithm named MLLSFE is proposed,which could be used to achieve target of effectively reducing attribute dimension of label-specific feature space.According to each label-specific feature space,this proposed algorithm is based on adaptive adjacency relationship between different samples and those ideas of dimensionality reduction controlled by pair-wise constraint,then maps original high-dimensional features to the corresponding low dimensional projection space through linear mode.In the end,through comparative analysis those experimental statistical results on a series of related data sets,the effectiveness of the algorithm MLLSFE could be proved.