A transfer-learning method for supervised outlier detection where the rare class is scarce in the target domain. Working in the Neyman-Pearson framework (minimize Type-II error subject to a cap on Type-I error), the paper extends it to Lipschitz surrogate losses and proposes a model-free meta-algorithm: a constrained optimization that jointly uses abundant source outlier samples and scarce target samples. It proves a target generalization bound in terms of the source and target sample sizes, the Rademacher complexity of the hypothesis class, and a transfer exponent measuring source-target relatedness, guaranteeing that an unrelated source is ignored so negative transfer is avoided. An implementable neural-network instantiation via a Lagrangian over tuning parameters outperforms transfer baselines on climate and financial outlier-detection data.
A transfer-learning method for supervised outlier detection, where the rare (abnormal) class has very few target samples. The work is set in the Neyman-Pearson classification framework, which minimiz…