Research On Hierarchical Classification Based On Label Distribution

As the extension of standard classification problems,hierarchical classification problems use the pre-defined label hierarchical structure to improve the classification performance,which is widely used in image classification,text classification and other fields.Over the past decade,hierarchical classification problems have attracted the attention of many researchers,and a large number of excellent hierarchical classification algorithms have been proposed.However,the existing algorithms lack two aspects of research on hierarchical classification problems: 1.There is small data issue in hierarchical classification;2.The characteristics of unordered and ordered label hierarchical classification problems are different,and special algorithms need to be designed according to their characteristics.For the first aspect,the small data issue is that the training samples of local modules are insufficient,and the closer to the bottom,the fewer samples.Previous works directly train the classifiers on the local module.Due to the small data issue,the local classifiers are prone to over-fitting,which becomes a major bottleneck of hierarchical classification.For the second aspect,in the classification problem,the label space can be divided into two types: unordered label space and ordered label space.Since the hierarchical structure is constructed on the label space,according to whether the label space is orderly,the hierarchical classification problem can be divided into unordered label hierarchical classification problem and ordered label hierarchical classification problem.These two problems have different characteristics,and it is necessary to design special algorithms for each problem.Previous hierarchical classification algorithms neglect the difference,resulting in the consequence that the algorithms can not adapt to the two problems.The goal of this paper is to research the above two aspects of the hierarchical classification problem and propose corresponding solutions.The main contributions of this paper include: 1.For the small data issue in hierarchical classification,this paper believes that the labels in the local module are correlated,and the degree of the correlation is variant in different local modules.According to the above observations,this paper proposes a hierarchical classification algorithm based on label distribution,whose key idea is to use the label distribution to represent the correlation between labels,so that both true label and its siblings can provide supervision information for the instance;2.For the difference between unordered label hierarchical classification problem and ordered label hierarchical classification problem,this paper designs a special hierarchical classification algorithm based on label distribution for each problem,so that the algorithm can adapt to the characteristic of the problem.This paper consists into six chapters.Chapter 1 mainly introduces the concepts and research status of hierarchical classification and label distribution learning,as well as the research content of this paper;Chapter 2 introduces label distribution learning in detail,as well as existing algorithms and evaluation methods;Chapter 3 introduces hierarchical classification in detail,as well as existing algorithms and evaluation methods;Chapter 4 introduces the method that uses label distribution to solve unordered label hierarchical classification problem.Chapter 5 introduces the method that uses label distribution to solve ordered label hierarchical classification problem;Chapter 6 summarizes the work of this paper.