Leveraging Neighbor Attributes for Classification in Sparsely Labeled Networks

摘要: Many analysis tasks involve linked nodes, such as people connected by friendship links. Research on link-based classification (LBC) has studied how to leverage these connections improve accuracy. Most prior research assumed the provision of a densely labeled training network. Instead, this article studies common and challenging case when LBC must use single sparsely network for both learning inference, where existing methods often yield poor To address challenge, we introduce novel method that enables prediction via “neighbor attributes,” which were briefly considered early work but then abandoned due perceived problems. We explain, using extensive experiments loss decomposition analysis, neighbor attributes significantly improves further show appropriate semi-supervised (SSL) is essential obtaining best accuracy in domain gains remain across range SSL choices data conditions. Finally, given challenges label sparsity impact attributes, multiple previous be re-considered, including regarding model features, noisy strategies active learning.