This library provides survival permanental process (SurvPP) implemented in Tensorflow. SurvPP provides a scalable Bayesian framework for survival analysis with time-varying covariates. For details, see our NeurIPS2023 paper [1].
The code was tested on Python 3.10.8, pandas 2.0.1, tensorflow-deps 2.10.0, tensorflow-macos 2.10.0, and tensorflow-metal 0.6.0.
To install latest version:
pip install git+https://github.com/HidKim/SurvPP
Import SurvPP class:
from HidKim_SurvPP import survival_permanental_process as SurvPP
Initialize SurvPP:
model = SurvPP(kernel='Gaussian', eq_kernel='RFM', eq_kernel_options={'n_rfm':500})
kernel: string, default='Gaussian'
The kernel function for Gaussian process. Only 'Gaussian' is available now.
eq_kernel: string, default='RFM'
The approach to constructing equivalent kernel. Only 'RFM' is available now.
eq_kernel_options: dict, default={'n_rfm':500}
The options for constructing equivalent kernel.
'n_rfm'specifies the number of feature samples for the random feature map approach to constructing equivalent kernel.
Fit SurvPP with data:
time = model.fit(formula, df, set_par, lr=0.05, display=True)
formula: string in the form 'Surv(Start, Stop, Event) ~ cov1 + cov2 + ...'
Identify the column labels of start time (Start), stop time (Stop), event indicator (Event, 1 represents that an event occurred at stop time, and 0 represents that observation is right-censored), and covariates used for survival analysis.
df: pandas.DataFrame
The survival data in counting process format. Each row represents a sub-interval for a subject, and should contain start and end time points, event indicator (0/1), and values of covariates in the sub-interval.
set_par: ndarray of shape (n_candidates, dim_hyperparameter)
The set of hyper-parameters for hyper-parameter optimization. The optimization is performed by maximizing the marginal likelihood.
lr: float, default=0.05
The learning rate for gradient descent algorithm (Adam).
display: bool, default=True
Display the summary of the data and the fitting.
- Return: float
The execution time.
Predict hazard function on specified covariate values:
r_est = model.predict(y, conf_int=[0.025,0.5,0.975])
y: ndarray of shape (n_points, dim_covariate)
The points on covariate domain for evaluating intensity values.
conf_int: ndarray of shape (n_quantiles,), default=[.025, .5, .975]
The quantiles for predicted hazard function.
- Return: ndarray of shape (n_quantiles, n_points)
The predicted values of hazard function at the specified points.
- Hideaki Kim. "Survival Permanental Processes for Survival Analysis with Time-Varying Covariates", Advances in Neural Information Processing Systems 36, 2023.
@inproceedings{kim2023survival,
title={Survival Permanental Processes for Survival Analysis with Time-Varying Covariates},
author={Kim, Hideaki},
booktitle={Advances in Neural Information Processing Systems 36},
year={2023}
}
Released under "SOFTWARE LICENSE AGREEMENT FOR EVALUATION". Be sure to read it.
Feel free to contact the author Hideaki Kim (hideaki.kin@ntt.com).