Redundancy reduction based node classification with attribute augmentation

Abstract Node classification for attributed graphs has attracted more and more researchers, it is very useful when labeled data are expensive and hard to obtain. However, most existing methods either only focus on node attribute data or are designed for working in single mode, resulting in omitting extra information and impairing classification performance. In view of this, in this work, we propose a semi-supervised method that can effectively incorporate various available prior information to augment the attribute matrix. This approach adds a two-level learning strategy to select and find those discriminative attributes for classification, and an augmentation step to combining various information, thereby reducing the redundancy exists between attributes and strengthening the community structure. It provides a unified way that can preserve both the node and edge information of the network. Finally, we use the classical classifier SVM to partition the induced augmented graph, we compare our method with 5 state-of-the-art methods on 4 data sets, the results confirm the effectiveness of our method.