FS-MGKC: Feature selection based on structural manifold learning with multi-granularity knowledge coordination

Abstract

Feature selection uses the hierarchical dependency information provided by multi-granularity knowledge to identify the most relevant features for a given task. Most of these methods manually assign parameters to adjust the restrictive relationships between these dependencies including strong and weak links. However, manual parameter tuning is challenging to accurately determine the influence degree of each parameter to harm performance, because the parameter tuning process involves the interdependence between the multiple hierarchical structures. Along these lines, we propose feature selection based on structural manifold learning with multi-granularity knowledge coordination to balance local structure relationships automatically. First, an affinity matrix based on multi-granularity knowledge is defined as a manifold coordination analysis, which learns all dependencies without destroying the original class space. This analysis effectively trade-offs these dependencies to adjust inter-class alienation and intra-classes inseparable. Afterward, feature correlation and sparsity are leveraged to select features with the most distinguishable and informative features as differentiation analysis. These two analyses jointly selected the feature subsets that are ask-related and informative in the original knowledge structure. Additionally, the proposed method has good universality, which can handle hierarchical classification tasks of the different structure types. Different hierarchical classification results demonstrate the effectiveness of the developed method.