© 2024Artificial immune detectors are the basic classification units for self/nonself discrimination. Traditional immune detector generation algorithms adopt supervised learning paradigm, relying on a large number of labeled samples to fully train the candidate detectors. However, in practical applications it is often difficult to obtain sufficient labeled training samples, resulting in model's insufficient learning problems. In the paper, we proposed a semi-supervised detector generation algorithm DE-PSA, which generates immune detectors using some initially labeled samples and a large number of unlabeled samples. DE-PSA consists of two main steps: dual-stage label propagation and detector generation. In the first stage of label propagation, pseudo labels are propagated to each unlabeled sample from its k-nearest labeled neighbors based on category influence calculations. According to the calculation results, we can select samples with highly credible pseudo (HCP) labels and partial labeled (PAL) samples which have multiple candidate labels. In the second label propagation stage, following partial label learning theory, category probabilities are iteratively propagated from the initially labeled and HCP labeled samples to the PAL samples to achieve label disambiguation; Subsequently, self (positive) and nonself (negative) samples are selected from the initially labeled samples, HCP labeled samples, and disambiguated PAL samples to constitute training set. Based on the set, DE-PSA generates self-detectors with variable radii using a positive selection process. Comprehensive tests on 10 standard datasets are carried out to test DE-PSA, and the true positive rates of self/nonself samples: TPS and TPN are taken as evaluation metrics. The results show that DE-PSA outperforms traditional algorithms, such that compared with newly proposed DGA-PSO, SA-PSA,HI-Detector, the average true positive rate of DE-PSA is raised by 23%, 19.5%, 16.5% respectively, and when compared with state-of-the-art algorithm co-PSA, only with 0.1‰ initially labeled training samples, DE-PSA and co-PSA has similar TPS, but DE-PSA's TPN is raised by 30%.
NETL
NSTL
Elsevier
