Censored Quantile Regression Neural Networks for Distribution-Free Survival Analysis

Authors: Tim Pearce, Jong-Hyeon Jeong, yichen jia, Jun Zhu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 6 the algorithm s effectiveness is empirically demonstrated by benchmarking against alternative methods on simulated and real data. Careful ablations illuminate effects of the algorithm s hyperparameters.
Researcher Affiliation Collaboration Tim Pearce1,2, , Jong-Hyeon Jeong3, Yichen Jia3, Jun Zhu1 1Dept. of Comp. Sci. & Tech., NRist Center, Tsinghua-Bosch Joint ML Center, Tsinghua University 2Microsoft Research, 3University of Pittsburgh
Pseudocode Yes Algorithm 1 Sequential grid algorithm for NNs. Algorithm 2 CQRNN algorithm.
Open Source Code Yes Code: https://github.com/Tea Pearce/Censored_Quantile_Regression_NN.
Open Datasets Yes Table 2: Summary of all datasets used. The paper lists standard benchmark datasets like Housing, Protein, Wine, PHM, Surv MNIST, METABRIC, WHAS, SUPPORT, GBSG, TMBImmuno, Breast MSK, LGGGBM, and provides detailed characteristics for them, implying public availability.
Dataset Splits No Table 2 specifies 'Train' and 'Test' data sizes, but no explicit 'validation' dataset split or sample counts are provided in the main text or tables.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, memory) used for experiments are provided in the paper.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) are listed in the paper.
Experiment Setup Yes All experiments use fully-connected NNs with two hidden layers of 100 neurons, except for Surv MNIST, when three convolutional layers are used. Grid size M is set to 5, 9 or 19 depending on dataset size...using cy = 1.2 provided consistently reasonable results... Appendix B.1: For all experiments we use the Adam optimizer with a learning rate of 1e-3, and a batch size of 256 for all datasets except those with N>20000, where a batch size of 512 is used. We train for 200 epochs unless stated otherwise.