Adversarial de-overlapping learning machines for supervised and semi-supervised learning

Yichen Sun, Chi-Man Vong, Shitong Wang

Abstract: While adversarial link information like the commonly used must-link and cannot-link constraints on training data are available, the existing AUC maximization learning frameworks cannot explicitly incorporate them to better guide disentanglements of the overlapping areas. As the first attempt in filling this gap, this study first develops the coupling-based adversarial overlapping concept by means of the coupling of the classical AUC with the modularity caused by adversarial link information. Then the corresponding adversarial de-overlapping maximization learning machine called De-OVL for supervised imbalanced data is developed. Furthermore, by using the proposed two-channel based strategy, De-OVL is extended to its semi-supervised version SDe-OVL with only one tunable hyperparameter for semi-supervised imbalanced data. Based on random Fourier features (RFF), the fast training versions RFF-De-OVL and RFF-SDe-OVL are developed to scale up De-OVL and SDe-OVL, respectively. In contrast to existing imbalanced classification methods, De-OVL has its unified adversarial de-overlapping maximization framework for supervised and semi-supervised imbalanced data, with fewer hyperparameters to be tuned. Extensive experimental results on four groups of benchmarking imbalanced datasets verify the above effectiveness of the proposed machines.