Research On Label Weighted Multi-label Feature Selection Algorithm

Nowadays,citizen's daily necessities,food,shelter,and transportation are all improving towards intelligence,and all of this is inseparable from the mining of data and information.Smart lifestyles generate huge amounts of data,and data mining is the process of discovering potential information from a large amount of that data.In the field of data mining,the dimensionality disaster caused by high dimensional data and data overfitting are two major problems that plague researchers.A new feature subset is formed by selecting important features from the candidate feature set,and using this subset for learning tasks such as model training.Feature selection can effectively overcome the dimensional catastrophe and reduce overfitting.With the emergence of multi-label problems in more and more fields,the research on multi-label feature selection algorithms has become more popular.Most of the existing multi-label researches are confined to a fixed thinking and only consider that the impact of each label on the feature selection model is the same.However,if the rich information of each label is used,the performance of the feature selection model can be improved.In this paper,two novel methods are proposed by making full use of implicit label information through label weighting.The main research work is as follows:(1)A algorithm based on fuzzy boundary weighted labels is proposed.First of all,this paper introduces fuzzy rough set for multi-label feature selection.The concept of fuzzy boundary domain is used as a criterion for evaluating the importance of tags and feature importance.The importance of tags is used to explore tag information to further assist feature selection.Then,this paper designs a dynamic label weighting heuristic algorithm to select features.Finally,experiments illustrate that the algorithm on the multi-label data set has better classification performance.(2)A multi-label I-Relief algorithm based on weighted labels with mutual information is proposed.First of all,this paper introduces I-Relief to tackle the trouble of multi-label,and iteratively updates feature weights for the purpose of maximizing the distinguishing ability of samples.Then,In the iterative process,label weights are added to strengthen the distinguishing ability of samples.Finally,this paper designs an algorithm to address the optimization case,and uses experiments to prove the effectiveness of all algorithms.