GRAFT: Decoupling Ranking and Calibration for Survival Analysis

2026-02-08

AI Summary

This work proposes GRAFT, a novel model addressing the challenges of censored data, high-dimensional features, and nonlinear interactions in survival analysis. GRAFT is the first method to decouple prognostic ranking from risk calibration within the accelerated failure time (AFT) framework, combining a linear AFT backbone with a nonlinear residual network. It introduces a stochastic gating mechanism for end-to-end sparse feature selection and incorporates two key innovations: a differentiable C-index alignment loss to optimize ranking performance and a stochastic conditional imputation strategy based on local Kaplan–Meier estimates to handle censoring. Extensive experiments demonstrate that GRAFT significantly outperforms existing approaches across multiple public benchmarks, achieving state-of-the-art performance in discrimination, calibration, and robustness to high noise with sparse features.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Graphical Models
• Search and Optimization: Metareasoning and Metaheuristics

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Survival analysis is complicated by censored data, high-dimensional features, and non-linear interactions. Classical models are interpretable but restrictive, while deep learning models are flexible but often non-interpretable and sensitive to noise. We propose GRAFT (Gated Residual Accelerated Failure Time), a novel AFT model that decouples prognostic ranking from calibration. GRAFT's hybrid architecture combines a linear AFT model with a non-linear residual neural network, and it also integrates stochastic gates for automatic, end-to-end feature selection. The model is trained by directly optimizing a differentiable, C-index-aligned ranking loss using stochastic conditional imputation from local Kaplan-Meier estimators. In public benchmarks, GRAFT outperforms baselines in discrimination and calibration, while remaining robust and sparse in high-noise settings.

Problem

• Survival Analysis
• Censored Data
• High-dimensional Features
• Model Calibration
• Interpretability

Innovation

• decoupled ranking and calibration
• gated residual network
• stochastic feature selection
• differentiable C-index loss
• survival analysis