Label-specific Features Learning With Complexity Reduction And Label Complement

With the advent of the 5G era,the amount of data has exploded.How to effectively process and analyze complex data has become the research focus of current scholars.Among many essential research branches,multi-label learning is fascinating.Different from traditional single-label learning,an instance is only associated with the single label.While each example is associated with multiple labels simultaneously in multi-label learning.Multi-label learning is more in line with the expression of ambiguous objects in the real world,so it is widely used in many fields.However,the traditional multi-label classification algorithms use the same feature set to predict all labels.In multi-label classification,a label may only be determined by its specific features,and these specific features are called label-specific features.It can effectively avoid certain useless features from affecting the classification performance and reduce multi-label learning complexity.Label-specific features learning is an embedded feature selection algorithm,which implements feature filtering during the learning process,and then the unique features of each label are obtained.However,multi-label data is currently characterized by high dimensionality and complexity and is accompanied by many missing labels,resulting in label-specific features learning more challenging.Therefore,the focus of this thesis will be on complexity reduction and missing-labels completion for attribute learning.The main research elements are as follows:(1)Features reduction based on label hierarchy.Most existing label-specific features methods treat all labels as the same granularity hierarchy,and ignore the possible multigranularity hierarchy among labels.To a certain extent,considering the specific features of each level of label can effectively reduce the complexity of label-specific features learning.Simultaneously,coarse-grained label prediction exacerbates fine-grained feature learning,while fine-grained features improve the learning of coarse-grained classifiers.To improve the classification performance and reduce the complexity of multi-labeling learning,a novel crossgranularity hierarchy for label-specific features learning is proposed by using the Huffman coding strategy to construct a multi-granularity hierarchy between labels.Firstly,the corresponding Huffman tree is constructed by calculating the frequency of each label among the positive instances.To realize the granularity hierarchy of labels,the hierarchical structure of each label is encoded by the Huffman tree.Then,the sparse linear model is used to obtain the local hierarchy-specific features.Finally,the extreme learning machine of the single-hidden layer feedforward neural network is used to predict labels of each hierarchy and form the final classification model.(2)Features reduction based on label correlation.The existing label-specific features methods only extract the important features from the label,while ignoring extracting important labels from the feature.It is easier to extract the unique features for labels by focusing on specific labels from the feature.Based on this,a novel label-specific features learning algorithm for multi-label classification is proposed.The algorithm combines the label's prominent features with the feature's essential labels.Firstly,to ensure the efficiency and accuracy of the model,the extreme learning machine is used to construct the joint model.Subsequently,the elastic network regularization theory is applied to the extreme learning machine's loss function.The correlation matrix of feature-specific labels is constructed as L2 term using mutual information theory,while the label-specific features are extracted by L1 term.(3)Recovery based on missing labels.Most existing label-specific features learning algorithms assume that label space is complete,ignoring the influence of missing labels on the classification accuracy.Some methods try to recover the missing labels first and then learn the mapping between the completed label matrix and the feature matrix.However,early intervention in the recovery of missing labels may affect the distribution of original labels to a certain extent.In this thesis,Feature-Label Dual-Mapping for missing-label-specific features learning is proposed.According to the information that the label depends on the feature,the dual mapping weight of the complete feature space and the missing label space is jointly learned.Therefore,the proposed algorithm is to conduct latent missing labels recovery by feature-label dual-mapping to directly obtain target weight in this thesis,avoiding the negative influence of early label recovery intervention.