Online Streaming Feature Selection for Class-Imbalanced Data Oriented Hierarchical Classification

In numerous practical domains, the feature space of data exhibits distinct characteristics, such as high dimensionality and dynamism, while the label space demonstrates a hierarchical structure and typically follows a long-tail distribution. To address this type of data, we propose an online streaming feature selection method based on hierarchical neighborhood consistency, which can select discriminative features from the streaming feature space. First, the number of neighbors for the target sample is determined by the number of samples in its class, and whether these neighbors are positive or negative is identified using the sibling strategy. Based on this novel hierarchical neighborhood relationship, hierarchical neighborhood consistency is defined at three levels: the individual sample, the layer, and the entire tree. Second, based on hierarchical neighborhood consistency, we establish three criteria for evaluating online streaming features: online importance analysis, online correlation analysis, and online redundancy analysis. Additionally, a framework for selecting online streaming features is designed. Finally, extensive experiments conducted on six hierarchical datasets demonstrate that the proposed algorithm significantly improves prediction accuracy for minority classes compared to several popular online streaming feature selection algorithms.