A Novel Semi-Supervised Classification Method Based on Soft Evidential Label Propagation

Abstract

For semi-supervised learning, we propose the label propagation method in the framework of evidence theory. The proposed method can deal with two problems which are the classification tasks with extremely few labeled data and identification of outlier samples. By incorporating the basic belief assignment functions as the soft evidential labels, both the uncertainty of samples' class assignments and the degree of samples' abnormality can be measured. On the basis of the designed label update mechanism, the soft evidential label of each initially unlabeled sample will be iteratively updated by absorbing the label information from its neighbors. Once the partition of the dataset is stable, the samples with high confidences of abnormality will be identified as outliers and others will be assigned to their most supported classes. The experiment results on several UCI and benchmark datasets validate the effectiveness of the proposed method in comparison with several state-of-the-art methods. In addition, tests on noisy spiral datasets show the robustness and feasibility of detecting outliers.