Effectively Self-Labeled Method with Accelerated Particle Swarm Optimization-Based Sample Subspace Optimization

Mislabeling is a challenging problem in self-labeled methods utilizing labeled and unlabeled samples to train classifiers. For overcoming misprediction, the existing self-labeled methods employ ensemble classifiers to improve predictive performance on unlabeled samples or employ heuristic strategies to find correctly predicted samples. Yet, they experience the following issues: (a) adopted ensemble classifiers employ random resampling or entire labeled samples without considering training sample quality and usefulness; (b) adopted heuristic strategies for selecting high-confidence pseudo-labeled samples explicitly rely on assumptions about geometric and class relationships; and (c) a few adopt ensemble classifiers and heuristic strategies to overcome misprediction. Inspired by the state-of-the-art sample subspace optimization with accelerated binary particle swarm optimization (SSO-ABPSO), an ensemble self-labeled method (ESLM-SSO) based on sample subspace optimization with accelerated binary particle swarm optimization is proposed against the above issues. The novelties of ESLM-SSO are as follows: (a) a SSO-ABPSO-based ensemble classifier considering training sample quality and usefulness is proposed to predict unlabeled samples, and (b) a SSO-ABPSO-based heuristic strategy that evaluates pseudo-labeled samples based on empirical subset performance, rather than explicit geometric assumptions, is proposed to identify correctly predicted samples. Experiments demonstrated that ESLM-SSO outperforms seven state-of-the-art self-labeled methods in improving two classifiers on extensive datasets.