Integrated self-supervised label propagation for label imbalanced sets

Label propagation is an essential graph-based semi-supervised learning algorithm. However, the algorithm has two problems: how to effectively measure sample similarity and handle label imbalanced sets. Recent advances in the semi-supervised field have shown that self-supervised pre-training can help to model more explicit class boundaries and significantly improve classification accuracy. Inspired by the problems and progress, we propose an integrated self-supervised label propagation algorithm (ISSLP). Our algorithm’s framework introduces self-supervised information in the iterative label propagation process to establish more explicit inter-class boundaries and control the label propagation scope. In addition, considering that different self-supervised models can capture different feature views of the data, we propose an entropy approach to measure the quality of feature views and generate weights for integrating multiple propagation results. The experimental results on synthetic and real-world datasets show that the proposed method can effectively solve the label imbalance problem and is insensitive to the parameters in the graph construction stage.