DOI: 10.11648/j.ajtas.20261505.16 ISSN: 2326-9006

Meta-Learning for Bayesian Computation in Survival Analysis: A Machine Learning Framework for Intelligent Bayesian Algorithm Recommendation

Khalid Salah, Afnan Salah
Show PDF Cite
Bayesian computation for survival analysis often involves selecting among alternative Markov chain Monte Carlo (MCMC) algorithms whose performance can vary substantially with the characteristics of the data and model. In practice, this selection is frequently based on trial and error, requiring repeated model fitting and computational assessment. We propose a meta-learning framework that formulates Bayesian computational strategy selection as a data-dependent prediction problem. A simulation-based meta-database is constructed by generating survival datasets under systematically varied sample sizes, model dimensions, predictor correlations, and censoring proportions. Four Bayesian computational strategies—Standard Random-Walk MCMC, Adaptive MCMC, Hamiltonian Monte Carlo, and Reversible Jump MCMC—are evaluated across the simulated environments. Estimation accuracy, sampling efficiency, and computational cost are integrated through a multi-criteria utility function to identify the preferred computational strategy for each dataset. Machine-learning models are then trained to learn the relationship between survival-data characteristics and the resulting algorithm-selection decisions. The learned recommendation models are evaluated using three independent real survival datasets: the German Breast Cancer Study Group breast cancer, Veteran lung cancer, and primary biliary cirrhosis datasets. Importantly, the real-data evaluation is conducted without retraining or recalibration of the meta-learning models. The proposed framework provides a systematic approach for transferring computational experience from simulated survival environments to new problems. By treating Bayesian algorithm selection as a data-dependent decision rather than assuming a universally optimal MCMC strategy, the framework offers a practical approach for reducing trial-and-error computation and supporting evidence-based selection of Bayesian computational methods.

More from our Archive