Selective Transfer Classification Learning With Classification-Error-Based Consensus Regularization

Transfer learning methods are conventionally conducted by utilizing abundant labeled data in the source domain to build an accurate classifier for target with scarce data. However, most current transfer assume that all relevant domain, which may induce negative effect when assumption becomes invalid as many practical scenarios. To tackle this issue, key is accurately and quickly select correlated corresponding weights. In paper, we make use of least square-support vector machine (LS-SVM) framework identifying their weights from domain. By keeping consistency between distributions classification errors both domains, first propose classification-error-based consensus regularization (CCR), can guarantee performance improvement classifier. Based on approach, a novel CCR-based selective method (CSTL) then developed autonomously choose exploit transferred knowledge solving LS-SVM based objective function. This minimizes leave-one-out cross-validation error despite training The advantages CSTL demonstrated evaluating its public image text datasets comparing it state-of-the-art methods.