Interactive Causal Correlation Space Reshape for Multi-Label Classification

Most existing multi-label classification models focus on distance metrics and feature spare strategies to extract specific features of labels. Those models use the cosine similarity to construct the label correlation matrix to constraint solution space, and then mine the latent semantic information of the label space. However, the label correlation matrix is usually directly added to the model, which ignores the interactive causality of the correlation between the labels. Considering the label-specific features based on the distance method merely may have the problem of distance measurement failure in the high-dimensional space, while based on the sparse weight matrix method may cause the problem that parameter is dependent on manual selection. Eventually, this leads to poor classifier performance. In addition, it is considered that logical labels cannot describe the importance of different labels and cannot fully express semantic information. Based on these, we propose an Interactive Causal Correlation Space Reshape for Multi-Label Classification (CCSRMC) algorithm. Firstly, the algorithm constructs the label propagation matrix using characteristic that similar instances can be linearly represented by each other. Secondly, label co-occurrence matrix is constructed by combining the conditional probability test method, which is based on the label propagation reshaping the label space to rich label semantics. Then the label co-occurrence matrix combines with the label correlation matrix to construct the label interactive causal correlation matrix to perform multi-label classification learning on the obtained numerical label matrix. Finally, the algorithm in this paper is compared with multiple advanced algorithms on multiple benchmark multi-label datasets. The results show that considering the interactive causal label correlation can reduce the redundant information in the model and improve the performance of the multi-label classifier.