The New England Journal of Statistics in Data Science

Severe event class imbalance is observed in many fields, such as insurance fraud, severe weather, traffic safety, and rare disease, and has 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 with limited samples, whose exceedance leads to extreme events. To accurately capture the extended tail, our solution adopts the Gumbel distribution as prior and is modeled with a variational inference framework. Following the assertion that exceedance leads to extremes, we devise a disentangled additive monotonic neural architecture to predict the risk. The proposed model acknowledges representation uncertainties while embracing 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 studies on various real-world data against the state-of-the-art counterparts, with encouraging results reported.