Research On Feature Selection Algorithm Based On Hierarchical Classification

In recent years,high-dimensional classification tasks have become an important challenge in machine learning.The difficulty of the classification task is increased by the highdimensional features and large number of categories in the data,and there are often hierarchical relationships between categories.Traditional planar feature selection algorithms do not consider the hierarchical structure of categories,resulting in unstable performance in the classification task.Although some hierarchical feature selection algorithms consider the hierarchical structure of categories,they lack stability when dealing with noisy data and do not exploit semantic relationships between categories,leading to limited exploration of hierarchical information.Given the inadequacies of existing feature selection algorithms,two hierarchical classification-based feature selection algorithms are proposed in this paper.The main research content includes two aspects:The problem of existing hierarchical feature selection algorithms not fully utilizing interlayer relationships and poor stability in the face of noisy data is addressed by decomposing the overall classification task into multiple subtasks.Based on the hierarchical relationship between classes,the similarity of features between parent classification tasks and sub-classification tasks is considered,a regularization term for interlayer parent-child relationships is added to the model,and a stable loss function is designed to filter out noisy data.A sparse regularization term is also introduced to ensure the sparsity and generalization ability of the model.Finally,a hierarchical classification feature selection algorithm based on interlayer parent-child relationships is proposed.The problem of insufficient utilization of inter-layer relationships and poor stability in the presence of noisy data in existing hierarchical classification feature selection algorithms is addressed by decomposing the entire task and considering the inclusion relationship between the current classification task and its descendant sub-classification tasks.An inter-layer inclusion constraint is defined in the model.A noise constraint mechanism is added to the loss function to reduce errors in the classification process,and a sparse regularization term is added to the model.Finally,a hierarchical classification feature selection algorithm based on interlayer inclusion constraints is proposed.In conclusion,two stable hierarchical classification feature selection algorithms are designed in this paper for hierarchical structure tasks based on inter-layer semantic relationships and noise constraints.The effectiveness of the proposed algorithms is compared with existing feature selection methods through experiments.The experimental results confirm the effectiveness of the proposed algorithms.