Workflow from Scientific Research

Framework of our proposed method. We first learn the source GCN and classifier using labeled source graph. In the process, the noisy nodes (contain noisy links or corrupted attributes, or both) are suppressed and their negative impacts are eliminated. Next, we learn the target GCN based on extending the Wasserstein distance to avoid noisy source nodes as well. At the testing time, the target nodes in the shared space are classified by the source classifier learned in the first step.

DOI