A novel ensemble label propagation with hierarchical weighting for semi-supervised learning

Nowadays, semi-supervised learning is one of the research sports to solve the problem of labeled data. Label propagation (LP) is a popular method to classify graph data by utilizing the characterization of data nodes. However, LP has randomness and cannot effectively employ node attributes. In this paper, to solve the above problem, we propose a novel LP method based on hierarchical weighting (HWLP). Firstly, attribute aggregation and attribute update are executed for each node. Secondly, in the process of LP, for each unlabeled node, the label owned by its neighbors with the highest similarity is selected to avoid arbitrary LP. Finally, maximum voting is adopted to enhance the stability of the results. Experimental results show that the proposed method has better performance and stability than others.