Research On Dynamic Knowledge Maintenance And Label Distribution Feature Selection

In the classic machine learning problem,multi-label learning as a popular research content is of great significance to the study of real-world ambiguous objects.However,the imbalance between labels are ignored in the existing multi-label learning algorithms.Many times,we are more concerned about the membership of the instance for different labels and their distribution,so there is research to replace the logical labels in the multilabel data with the form of the probability distribution,which gives the label distribution data.Whether it is multi-label data or label distribution data,it faces the problem of explosive growth of data size.Today,there are many feature selection methods for multilabel data,but there are not many researches on feature selection that can be applied to label distribution data.In fact,feature selection can help us discover the potentially important features and key information contained in the label distribution data.In addition,it can reduce the time and complexity of label distribution learning.Therefore,on the one hand,this paper proposes two kinds of feature selection algorithms suitable for conditional probability form label values in label distribution data to simplify complex and redundant label distribution data and facilitate label distribution learning.On the other hand,this paper proposes a corresponding dynamic knowledge maintenance method for the complex situation of system decision attribute value changes,in order to better maintain the key information and knowledge in the data.This paper focuses on dynamic knowledge maintenance and feature selection of label distribution data.The main research work and contributions are as follows:(1)This paper propose two label distribution feature selection algorithms based on fuzzy rough set model.One of the algorithms measures the correlation between features and label distribution and the redundancy between features based on fuzzy mutual information.Finally,the heuristic forward search strategy is adopted to select the feature subset.Another algorithm converts the fuzzy similarity for the label distribution data into a classical equivalence relation based on the fuzzy cut-off relationship.The fuzzy generalized decision is used to generate the assignment difference matrix and the minimum difference attribute set.Finally,the feature reduction output feature subset is performed.Theoretical analysis proves the rationality of the algorithm,and the experimental results on the real data set further verify the effectiveness of the algorithm.(2)This paper present an incremental algorithm for dynamic maintenance decision rules for decision attribute value changes.Based on the effect of coarsening and refining of decision attribute value on the generalized decision of original data,this paper proposes the update rule for the assignment difference matrix in the complex case of decision attribute value change.Finally,the computational complexity of the theoretical analysis proves the rationality of these algorithms,and compared with the classical rule induction method on the real data set,the experimental results verify the effectiveness of these proposed incremental algorithms.