Consensus rate-based label propagation for semi-supervised classification

Abstract Label propagation is one of the most widely used semi-supervised classification methods. It utilizes neighborhood structures of observations to apply the smoothness assumption, which describes that observations close to each other are more likely to share a label. However, a single neighborhood structure cannot appropriately reflect intrinsic data structures, and hence, existing label propagation methods can fail to achieve superior performance. To overcome these limitations, we propose a label propagation algorithm based on consensus rates that are calculated by summarizing multiple clustering solutions to incorporate various properties of the data. Thus, the proposed algorithm can effectively reflect the intrinsic data structures, and yield accurate classification results. Experiments are conducted on various benchmark datasets to examine the properties of the proposed algorithm, and to compare it with the existing label propagation methods. The experimental results confirm that the proposed label propagation algorithm demonstrated superior performance compared to the existing methods.