Research On Label Distribution Learning Based On Label Morphological Fuzzines

Label distribution learning is a novel machine learning paradigm for tackling ambiguity classification problems.This paper observes some correlation between label distribution morphological ambiguity and sample ambiguity,and no previous work on label distribution learning has explicitly investigated the ambiguity on label morphology.In order to improve the prediction of label distribution learning on ambiguity classification problems,this paper investigates this label morphological ambiguity.Firstly,this paper introduces the concept of label morphological ambiguity in label distribution learning,gives a measure based on kurtosis,and experimentally explores the effect of different label morphological ambiguity samples on classification learning.Specifically,this paper conceptually discusses the association between label ambiguity and morphological ambiguity,i.e.,essentially label distribution morphological ambiguity is a mixture of label ambiguity and other factors.Then a suitable quantification of label morphological ambiguity is given,using the fourth-order central moments of the distribution,i.e.,kurtosis,to reflect the dispersion in the distribution.Subsequently,based on this quantification study,some scenarios of application of morphological ambiguity in practical problems are listed.In the study,this paper finds that data with low morphological ambiguity are usually more conducive to the learning of classifiers in ambiguity classification tasks,and controlled experiments are designed to verify this finding.During the research,this paper finds that data with low morphological ambiguity are usually more beneficial for classifiers to learn in ambiguity classification tasks,and a controlled experiment is designed to verify this finding.Based on the finding that low ambiguity samples are more favorable for classification learning,this paper designs an ambiguity-sensitive label distribution learning algorithm LDL-KQA,which employs a low ambiguity weighting strategy and an ambiguity alignment strategy,combined with the idea of curriculum learning,to force the classifier to make predictions corresponding to the ambiguity of sample labeling morphology by constraining the ambiguity loss.Ablation experiments verify the performance of the proposed algorithm,and the experimental results show that the proposed LDL-KQA algorithm can effectively improve classification accuracy.Secondly,this paper proposes a component enhancement framework CEMG for label distribution learning inspired by age distribution learning,using morphological ambiguity and Gaussian distribution as a priori information.As a novel learning framework,the label component enhancement framework can effectively exploit the hidden prior information in label distribution,and compensate for the small data set in label distribution learning by learning adjacent component scores.More specifically,the label component enhancement framework obtains prior knowledge about the degree of label morphological ambiguity,and augments each component of the distribution with a Gaussian prior,then trains a separate component learner for each component and predicts each component separately,and finally uses a combination strategy to aggregate the prediction results of these component quantifiers together to form the final distribution prediction.In order to solve the problem that the component model is difficult to fit on some components,this paper also conducts an in-depth study on the boosting technique and designs an adaptive boosting algorithm Adaboost-LDL dedicated to label distribution learning.In this paper,each module of the component enhancement model is designed and studied separately,and the main design is illustrated with rich experiments.The experimental results show that the proposed CEMG framework can effectively improve the basic label distribution learning algorithm.In addition,this paper also proposes a novel LDL prediction paradigm called mass point prediction.Mass point is a concept in component prediction to represent points that are clustered in probability density.Mass point prediction gives a prediction that contains richer information about the ambiguity compared to the traditional distribution prediction,which can provide a better reference in certain ambiguous classification scenarios.