Using cluster information to improve label propagation

Abstract

In graph-based semi-supervised learning, using few samples as supervised information is usually not enough for classification tasks, while more labeled samples are often challenging and time-consuming to obtain. In this study, we use the clustering results as prior knowledge to guide the classification process in graph-based learning. At first, we combine density peaks clustering and label propagation algorithm to obtain the cluster information. Subsequently, the cluster information is transformed into a style factor represented by a symmetric nonnegative matrix. At last, the labels of labeled objects are propagated to others using the style factor as the supervised information. We validate our method in real datasets, and the results show that our method has statistically improved the accuracy of classification.