Label enhancement via manifold approximation and projection with graph convolutional network

Chao Tan, Sheng Chen, Xin Geng, Yunyao Zhou, Genlin Ji

Abstract: Highlights•Propose two algorithms, the label enhancement based on feature representation(LEFR) and the LE based on graph convolutional network(LE-GCN).•The LEFR first apply manifold learning to map the relatedness between low-dimensional feature samples to the label space. The implicit correlation between low-dimensional feature and label spaces promotes the LE process.•GCN is extended to the LE field for LE-GCN to fully exploit the hidden relationships between nodes and labels. The label learning accuracy is improved by enhancing node information with threshold connections and label connections.