Reproduce “SSH-Net: A Deep Neural Network for Predicting Failure Time Distribution Functions under Competing Risks with Application to GPU Data” on OpenResearch

About this paper

Jie Min, Yueyao Wang, Mengkun Chen

Researchers at the University of South Florida, Zhejiang Gongshang University, and Virginia Tech developed SSH-Net, a deep neural network that predicts failure time distribution functions under competing risks by aligning its architecture with hierarchical data structures and employing piecewise constant hazard functions. The model achieved superior prediction accuracy, Brier scores, and AUC values compared to existing deep learning methods like DeepHit and NFG in simulations and on GPU reliability data.

Key insights

• Explicitly linking neural network architecture to the hierarchical structure of input data enhances interpretability and provides a principled, domain-knowledge-driven approach to hyperparameter selection.
• Modeling continuous time with piecewise constant hazard functions, combined with a smoothness penalty, enables flexible yet stable estimation of hazard dynamics without significant information loss.
• Designing distinct sub-networks for different covariate groups allows SSH-Net to better exploit the inherent structural information within complex data, leading to more accurate and robust predictions.

Summary by alphaXiv.