Uncertain multi-conceptual information acquisition and fusion for hierarchical classification

Abstract

Recently, hierarchical classification has become a hot research problem due to the wide existence of hierarchically structured data in the real world. However, some existing studies on hierarchical classification proposed a series of feature selection methods without considering the design of a stopping mechanism, while others designed the stopping mechanism without taking the feature selection into account. Despite the simplicity of integrating the above two parts, they are usually not compatible with each other, which lead to reduced performance of hierarchical classification models. In this work, we obtain a novel hierarchical classifier by constructing a hierarchical fuzzy concept-cognitive learning model (HFCCLM), in which incremental hierarchical feature selection is realized by the update of inclusion degree of fuzzy concepts, and a stopping mechanism is designed by learning uncertainties of samples matching nodes in a hierarchical tree structure. That is, feature selection and design of a stopping mechanism can be unified in the fuzzy concept-cognitive learning framework. Furthermore, by integrating the uncertainties of set approximation of clue based fuzzy reasoning and clue based fuzzy concept-cognitive learning for stopping samples at appropriate nodes, it significantly enhances the performance of the proposed hierarchical classifier. Finally, experiments are carried out to show the effectiveness of our model in achieving hierarchical classification tasks.