| Hierarchical classification learning aims to process large-scale data by exploiting the hierarchical relationships between data categories.The high dimensionality of the data feature space is one of the main challenges of this research.Compared to traditional feature selection methods,hierarchical feature selection methods select different feature subsets for each subtask of hierarchical classification,which can more effectively improve the training speed and prediction accuracy of classification models.However,the prevalence of fine-grained sparsity,class imbalance,and semantic gaps in hierarchical data greatly affects the performance of existing hierarchical feature selection algorithms.In order to explore the method of hierarchical feature selection in complex environments and select better feature subsets for hierarchical classification,this thesis conducts targeted research on the above-mentioned issues,and the main work includes:(1)For the problem of insufficient granularity of labels,mining fine-grained label information is a viable idea.In this thesis,we propose a Hierarchical Feature Selection from Coarse to Fine(HFSCF)algorithm,which can address the challenges caused by the scarcity of fine-grained labels in the hierarchical feature selection training process.Subject to the principle of sibling inconsistency,we use an embedded model to recursively select a subset of features and predict the missing fine-grained markers by a linear model during feature selection.In addition,a manifold regularization is embedded in the model to enhance the reliability of the linear model output labels.Experiments show that the HFSCF algorithm can perform the feature selection task excellently under different label granularity scenarios.(2)For the class imbalance problem,we consider training the current class by samples from the relevant class.In this thesis,we propose a Hierarchical Feature Selection based on Label Distribution Learning(HFSLDL)algorithm.The algorithm considers sample similarity and the number of shared ancestor nodes to evaluate the degree of similarity between categories and represents this degree by a label distribution paradigm.In addition,we consider the joint effect of common and label-specific features on hierarchical classification,and thus construct regular terms to learn subsets of features with greater discriminative ability.The results on several datasets with different degrees of imbalance show that the HFSLDL algorithm has significant advantages over the state-of-the-art hierarchical feature selection algorithm and can effectively alleviate the negative effects of class imbalance.(3)To address the semantic gap problem,we propose a Hierarchical Feature Selection with Label Enhancement(HFSLE)algorithm.This algorithm embeds label enhancement into the hierarchical feature selection process.This algorithm embeds label enhancement into the hierarchical feature selection process.It can enhance the logical labels to the label distribution that are required for the hierarchical feature selection training process,so as to obtain clear semantic information.In addition,we select features in the hierarchical subtasks that reflect the commonality within classes to match the feature subsets with the hierarchical semantic structure.This can alleviate the propagation of errors caused by the semantic gap in hierarchical classification.Compared with advanced hierarchical feature selection algorithms show that the HFSLE algorithm ranks first on all experimental datasets.Compared with the three methods proposed in this thesis,the HFSLE algorithm has a strong competitive performance. |
