Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees

Authors: Weijia Zhang, Chun Kai Ling, Xuanhui Zhang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments results from a wide range of datasets demonstrate that our approach successfully discerns the underlying dependency structure and significantly reduces survival estimation bias when compared to existing methods.
Researcher Affiliation Academia 1 School of Information and Physical Sciences, The University of Newcastle, Australia 2 Carnegie Mellon University, USA 3 School of Information Management, Nanjing University, China
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes 1https://github.com/Weijia Zhang24/DCSurvival
Open Datasets Yes We use two semi-synthetic datasets based on the STEEL (V E, Shin, and Cho 2020) which contains 35,040 samples with 9 covariates, and the Airfoil (Thomas Brooks 1989) datasets which includes 1503 samples with 6 covariates. ... The SEER dataset is from the Surveillance, Epidemiology and End Results database (Howlader et al. 2010). ... GBSG2 is from the German Breast Cancer Study Group (Schumacher et al. 1994)
Dataset Splits Yes We use Adam W (Loshchilov and Hutter 2019) for optimization and use 50%/30%/20% training/validation/test splits.
Hardware Specification Yes Experiments are conducted with one NVIDIA RTX4090 GPU.
Software Dependencies No We utilize Pytorch (Paszke et al. 2017) for implementing all neural networks and automatic differentiation.
Experiment Setup Yes Tensors are computed with double precision (fp64) as the inversion of φ mandates numerical precision. When using Newton s method to compute the inverse φ 1 nn, we terminate when the error is less than 1 10 12. For all our experiments we set φnn with L = 2 and H1 = H2 = 10, i.e., the copula representation contains two hidden layers with each of width 10. ... We use Adam W (Loshchilov and Hutter 2019) for optimization and use 50%/30%/20% training/validation/test splits. We used validation samples for early stopping based on the validation log-likelihood.