Research On Feature Optimization Method For Multi-Label Decision System

The realistic objects can be abstract into data samples,and the increasing dimension of data samples leads to the increasing complexity of information expression.On the other hand,realistic objects are often polysemous.Therefore,the concept of multi-label feature selection has become one of the research and application hotspots in data mining and pattern recognition,which aims to eliminate irrelevant features and redundant features,thereby improving learning efficiency and optimizing model performance.In this thesis,the UN-MLFSPO algorithm is proposed,which utilizes the Hilbert-Schmidt Independence Criterion(HSIC)to measure the degree of correlation between features and labels.And then,an approximation optimal feature subset is obtained in the expanded high-dimensional space by employing the Pareto optimal principle.Because of the need to set the number of feature subset manually,the MLFSPO algorithm is presented,further adopts the Pareto optimal principle to rank features,and introduces the label weight to obtain the final feature importance ranking.The experiments show the effectiveness of two multi-label feature selection algorithms.Nowadays,in practical applications,a large amount of data can be obtained,but information labeling is expensive.Therefore,in this thesis,a novel semi-supervised multi-label feature selection algorithm is proposed,called SMLFSLS.It defines the metric to calculate sample similarity,which utilizes the label information and geometric distribution information of features.The experiments show the effectiveness of the algorithm in semi-supervised multi-label learning.