Gradient boosting for time-to-event analysis
Contact person: Riccardo De Bin
Keywords: gradient boosting; first-hitting-time models; informative censoring; structural missingness; survival analysis
Research group: Statistics and Data Science
Department of Mathematics
Gradient boosting is a powerful machine-learning approach with the interpretability of a statistical model. Originally developed as a black-box for classification, it has been statistically interpreted and applied to many statistical problems. Technical advantages of boosting include the ability to cope with high-dimensional data, implicit variable selection and shrinkage. Here the idea is to focus on time-to-event analysis applications, where the response is the time to an event of interest. In this context, the idea of boosting has been successfully applied to standard problems, but there is a need for further developments that can handle more challenging situations, such as informative censoring and structural data missingness (e.g., blocks of data completely not available for specific groups of patients). In addition, structural issues of boosting, such as the choice of the value of the tuning parameter (number of boosting iterations) should be investigated specifically for the time-to-event analysis case.
Methodological research topics:
- Provide a unified framework for the application of boosting FHT models in time-to-event analysis;
- Develop boosting models to deal with informative censoring;
- Develop boosting algorithms to deal with structural missingness, with focus on time-to-event analysis;
- Study the stopping criteria to control the bias-variance trade-off and provide specific solutions for time-to-event data applications.
Mentoring and internship will be offered by a relevant external partner