FastSurvival: Hidden Computational Blessings in Training Cox Proportional Hazards Models

Authors: Jiachang Liu, Rui Zhang, Cynthia Rudin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we verify the computational efficiency of our methods. As a direct application, we show how our optimization methods can be used to solve the cardinality-constrained CPH problem, producing very sparse high-quality models that were not previously practical to construct. ... We test the effectiveness of our optimization methods on both synthetic and real-world datasets.
Researcher Affiliation Academia Jiachang Liu1, Rui Zhang2, Cynthia Rudin2 1Cornell University, 2Duke University
Pseudocode No The paper describes its methods verbally in text and mathematical derivations in the appendix, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Implementations of Fast Survival discussed in this paper are available at https://github.com/jiachangliu/Fast Survival.
Open Datasets Yes Flchain: We use the dataset from the Scikit-Survival [56] package. The Git Hub link to this dataset is https://github.com/sebp/scikit-survival/tree/master/sksurv/datasets/data. The license of this package is GPL-3. ... We have a summary of datasets for experiments in Table 1.
Dataset Splits Yes We perform 5-fold cross validation and report the mean and standard deviation of different metrics on both the training and test sets. ... We ran 5-fold cross-validation (random seed 0) on the following datasets: Dialysis, Flchain, Kickstarter1, Employee Attrition, Synthetic High Corr High Dim1, Synthetic High Corr High Dim2, Synthetic High Corr High Dim3.
Hardware Specification Yes All experiments were run on the Intel(R) Xeon(R) CPU E5-2680 v3 Processor, 2.50GHz. We set the memory limit to be 100GB.
Software Dependencies Yes Abess: ... We used the Cox model in abess python package Version 0.4.6. ... Sksurv Coxnet: ... scikit-survival version-0.20.0 (https: //scikit-survival.readthedocs.io/en/stable/index.html).
Experiment Setup Yes For each dataset, we ran algorithms with different configurations and evaluated fitted models with metrics described in Appendix C.2: Abess: We ran this algorithm with 30 different configurations: support size, k, ranging from 1 to 30, forcing the number of non-zero coefficients in the Cox model to be exact k. We set primary_model_fit_max_iter to be 20, approximate_Newton to be False. All other parameters were set to the default.