Multi granularity based label propagation with active learning for semi-supervised classification
Shengdan Hu,Duoqian Miao,Witold Pedrycz
DOI: https://doi.org/10.1016/j.eswa.2021.116276
IF: 8.5
2022-01-01
Expert Systems with Applications
Abstract:Semi-supervised learning (SSL) methods, which exploit both the labeled and unlabeled data, have attracted a lot of attention. One of the major categories of SSL methods, graph-based semi-supervised learning (GBSSL) learns labels of unlabeled data on an adjacency graph, where neighborhood sparse graph is often used to reduce computational complexity. However, the neighborhood size is difficult to set. Instead of assigning a concrete value of neighborhood size, we propose a new label propagation algorithm called multi granularity based label propagation (MGLP) and developed from the view of granular computing. In MGLP, labels of unlabeled data are learned by two classic label propagation processes with diverse neighborhood size k, where granular computing delivers a guiding strategy to leverage multiple level neighborhood information granules, and three-way decision acts as an active learning strategy to select the unlabeled data for further annotating. Through the iterative procedures of label propagating, data annotating and data subset updating, the ultimate pseudo label accuracy of unlabeled data may be higher. Theoretically, the accuracy of pseudo labels is enhanced in some scenarios. Experimentally, the results of simulation studies on ten benchmark datasets, show that the proposed method MGLP can rise pseudo labels accuracy by 8.6% than LP (label propagation), 6.5% than LNP (linear neighborhood propagation), 6.4% than LPSN (label propagation through sparse neighborhood), 4.5% than Adaptive-NP (adaptive neighborhood propagation) and 4.6% than CRLP (consensus rate-based label propagation). It also provides a novel way to annotate data.