Combine labeled and unlabeled data for immune detector training with label propagation

Artificial immune detectors are the basic recognition components of immune systems. Traditionally, the candidate non-self detectors are compared with the whole self training set to eliminate self-reactive ones in negative selection algorithms (NSAs). However, the training process has low efficiency due to the exhausting comparisons. Furthermore, it can be more efficient if we straightforwardly generate self-detectors based on the available self samples to avoid the overwhelmed comparisons. In the paper, a new detector training algorithm is proposed. Firstly, the self training set is enlarged by the label propagation algorithm (LPA) using both labeled and unlabeled samples; and then the newly labeled samples is evaluated based on noisy learning theory to remove the unqualified ones. Finally self-detectors are directly generated at the locations of self samples. The theoretical analysis demonstrated that the time complexity of our algorithm is much reduced, especially that the exponential relationship between self size and time complexity in traditional NSAs is eliminated. The experimental results showed that: not only the time cost of detector training, but also the detection accuracy is improved.