Mixup for Survival Analysis

Byeongchan Ryu, Chihyeon Choi, Youngdoo Son, Hyungrok Do

Keywords: mixup, survival analysis, time-to-event data, vicinal risk minimization, vrm

TL;DR: We propose a mixup strategy for survival analysis.

Abstract: Survival analysis with censored outcomes underpins risk prediction in domains such as medicine and engineering. While deep survival models capture complex input-output relationships, they may benefit from vicinal risk minimization techniques such as mixup, which have proven effective as simple, model-agnostic data augmentation methods. We introduce a mixup for survival analysis, called H-Mixup, a principled framework that adapts mixup to censored time-to-event data by defining augmentation strategies that yield interpolated outcomes consistent with valid survival trajectories, encouraging local linearity in hazard functions. Theoretically, we show that H-Mixup contracts empirical Rademacher complexity and can tighten generalization bounds, while noting that the overall bound still depends on vicinal bias, which varies with alignment between mixing assumptions and the underlying risk structure. Empirical results on semi-synthetic and real-world datasets suggest that H-Mixup improves predictive performance for deep survival models. Overall, this study addresses the gap between mixup and survival analysis, providing a general recipe for vicinal regularization via simple data augmentation in time-to-event modeling.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 9933