Anchor-based graph embedding and soft label learning for multi-label classification with missing label

Despite progress in multi-label learning, handling missing labels in weakly annotated data remains challenging. The graph-based method utilizes sample structure information to promote semantic reconstruction to recover missing labels, and it has achieved satisfactory performance. However, the existing graph-based soft label semantic reconstruction methods still have the following problems with missing label classification learning. Existing methods face two limitations: (a) the separate construction of instance similarity graphs and execution of classification tasks results in suboptimal similarity matrices for classification; (b) the high computational cost incurred by some adaptive $k$-connected graph learning methods when constructing $k$-nearest neighbor similarity matrices hinders their practical application. In addition, the complex label correlations further increase the difficulty of accurately predicting all possible labels. In this paper, we propose a joint bipartite graph and soft label learning for the multi-label missing label classification method. First, we obtain the anchor of the samples based on $k$-means clustering and construct a sparse bipartite graph similarity matrix with $k$-nearest neighbor connected components by using the anchor and the original samples as the weight of the bipartite graph. Second, sparse high-rank and high-order label correlation learning based on landmark selection is provided to obtain soft labels for the missing labels. Finally, learn multi-label classification and label correlations jointly in a unified framework with kernel techniques to solve the linear inseparability of the data. Experimental results on several real-world benchmark multi-label datasets demonstrate the competitiveness and effectiveness of our proposed method compared with the state-of-the-art approach.