Multi-label Feature Selection Based On Gravitational Field Model

Multi-label learning is widely used in machine learning and artificial intelligence.In the multi-label learning framework,it is necessary to collect a large amount of feature data to describe the object more accurately,but as the feature data increases,the redundant features also increase which directly affects the accuracy of the classifier and increases the model training time.Feature selection is an effective method for dealing with data high-dimensionality problems by deleting redundant or unrelated features in the original feature space to select a set of feature subsets containing all or most of the information of the original feature space.At present,most feature selection algorithms are basically based on the criterion of “maximum correlation and minimum redundancy”,but possible interaction between the features is ignored.In order to consider the interaction force between features,in this paper,we apply the repulsion loss and universal gravitation to the multi-label feature selection,each feature is treated as an atom,gravitation or repulsion also exists between the features,and treat these forces as forces in the gravitational field.In addition,the above method selection feature subsets are all obtained in advance based on the entire feature space before feature selection.However,in the actual situation,some problems are characterized by incremental or dynamic feature space and label space.How to deal with the streaming feature selection in this dynamic environment is worthy of study.In view of the above problems,the main research work of this paper is as follows:(1)In order to consider the force between features,the principle of mutual attraction and attraction of physical magnetic poles is used.Assuming that there is attraction or repulsive force between features,and a multi-label Feature Selection Algorithm based on Feature Repulsion Loss(FS-FRL)is proposed.Firstly,a feature attraction term(ATC)is constructed based on neighborhood information entropy to express the degree of the attraction about the candidate features to the label space.Then,two feature exclusion terms are defined,which is to measure the repulsion degree of candidate features to the selected feature subset(RFS)and the repulsion /attraction degree to the selected single feature(RFF).Finally,combine the feature attraction term and the feature repulsion to construct the function of the feature repulsion loss.(2)Aim at the problem of online stream feature selection,Feature Repulsion Loss for Streaming Feature Selection with Dynamic Sliding Window(SF-DSW-FRL)is proposed based on the feature repulsion loss model.The algorithm uses the sliding window mechanism of finite size to prejudge the convective features,selects the candidate feature set that is strongly related to the label space through the design criteria,and measures the importance of the feature by using the feature rejection loss function.(3)The information entropy metric method of neighborhood model magnifies the dependency degree of neighborhood radius on feature selection to some extent.So,in this part,we apply the neighborhood information entropy model and the universal gravitational law to multi-label feature selection,and proposes Multi-label Feature Selection Algorithm Based on Neighborhood Data to Gravity Model(FS-DGM).Using the property that gravitation is inversely proportional distance so the dependence of neighborhood radius(distance)on the result of feature selection is reduced.