Hierarchical Classification Based On The Class Hierarchical Granulation

Large-scale sample data,a large number of categories,and high-dimensional features en-rich the diversity of machine learning methods to solve problems.Humans usually organize data into a hierarchical structure in a multi-level granular manner of“coarse to fine”granularity to realize memory and retrieval.In addition,there are complex structural relationships among categories.This type of data structure provides new ideas for machine learning classification,but there are certain limitations:(1)The hierarchical classification task is affected by the inher-ent inter-level error propagation problem of the category hierarchy knowledge,which reduces the classification effect;(2)The problem of“semantic gap”that may exist in the hierarchical structure constructed based on semantic knowledge;(3)The imbalance of category distribution in real data breaks the basic assumption of the traditional categories'uniform distribution.This causes the previous hierarchical classification methods to perform poorly in large-scale classifi-cation tasks.In this thesis,we fully excavate the relationship between the original data and the category hierarchical structure knowledge for the hierarchical classification task.We design the hierarchical classification method based on hierarchical granulation.The main contents are as follows:(1)Hierarchical classification method based on knowledge-driven hierarchical gran-ulation.Traditional hierarchical classification methods have inherent inter-level error propa-gation problems.We adopt a multi-path prediction strategy for knowledge-driven.This method uses the hierarchical structure contained in the given semantic space to consider the relation-ship between the upper and lower granular layers and within each granular layer to realize the prediction from coarse to fine granularity.Finally,a hierarchical classification method based on knowledge-driven hierarchical granulation is proposed.(2)Hierarchical classification method based on data-driven hierarchical granulation.There is a semantic gap problem between the traditional semantic hierarchy and the category hierarchy embodied by data characteristics.Firstly,the scheme considers the implementation of data-driven hierarchical granulation from the perspective of original data.The main idea is to perform bottom-up granular clustering based on the similarity between different types of granular features.Secondly,we use _1,2-norm to capture the shared and specific features of each granularity layer.This operation achieves the effect of feature dimensionality reduction.Finally,a hierarchical classification method based on data-driven multi-granularity clustering is constructed.(3)Hierarchical classification method based on knowledge-driven and data-driven hierarchical granulation.Some real datasets are in imbalanced distribution of samples,which can not be applied to traditional hierarchical classification methods.We use the complemen-tarity of hierarchical structure knowledge and raw data to apply two-way driven strategy.The problem is localized and processed separately according to the distribution of the original data.Data-driven refers to using the similarity between data samples to“fine to coarse”granulation to build a hierarchical structure suitable for small samples.Then rely on the threshold control strategy for knowledge-driven realization of top-down classification of different parts.Final-ly,a hierarchical classification method based on knowledge-driven and data-driven hierarchical granulation is proposed.