Incremental label propagation for data sets with imbalanced labels

Yaoxing Li, Liang Bai, Zhuomin Liang, Hangyuan Du

Abstract: Label Propagation (LP) is a popular graph-based semi-supervised learning algorithm. However, the imbalanced label distribution of labeled data affects the performance of LP, which makes it prefer to allocate more points for the class with majority labels. To get rid of this deficiency, we propose an Incremental Label Propagation (ILP) framework. The ILP framework includes an incremental balance strategy and a multiple results integration method to reduce the uncertainty of balancing labeled data. The incremental balance strategy can gradually add trusted pseudo labels to balance the label distribution so as to reduce the uncertainty caused by pseudo labels. Then, the integration method evaluates the quality of the balanced propagation results obtained from the incremental balance strategy and combines them with different weights to generate a robust outcome. The experimental results show that the proposed framework can combine with different LP methods and effectively solve the performance degradation of the LP algorithm caused by imbalanced labels.