|
Abstract:
|
Severe event class imbalance poses a major challenge for real-world research applications, such as insurance fraud, severe weather, traffic safety, and rare disease, having far-reaching significance when the generalization of minority event classes is of primary interest. While existing solutions mostly focus on differential sampling or sample re-weighting approaches to alleviate the imbalance issue, we take a novel alternative view on promoting generalization. Our proposal is formulated under the generative Bayesian framework, positing that predictors are the stochastic proxies of latent causes, whose exceedance leads to extreme events. To accurately capture the extended tail, our solution adopts generalized Pareto distribution as prior and is modeled with variational inference. Our model acknowledges representation uncertainties while at the same time embraces improved interpretability, generalization, and robustness. We provide theoretical insights to show the merits of the proposed approach. To verify the effectiveness in empirical settings, we conducted extensive studies on various real-world benchmarks, with encouraging results reported.
|
