Skip to Main content Skip to Navigation
Master Thesis

Sélection de variables à l’aide de forêts aléatoires pour données de survie de grande dimension

Abstract : Today, the amount of information gathered at any given moment has increased exponentially with the advancement of science. As the databases available are increasingly large, their analysis can be problematic because not all existing methods are suitable for high-dimensional data. In addition, if we are interested in this type of survival data and we seek to explain which variables are the most predictive or associated with the occurrence of the event, a problem of variable selection arises and to date, few solutions exist to deal with it. The objective of this internship is first of all, to do a review on the existing methods of variable selection for high-dimensional survival data, then to propose an adaptation of a variable selection method with random survival forests and finally, to study its behavior and performance. The method of variable selection VSURF (Variable Selection Using Random Forests) developed by Genuer et al. in 2010 is adapted to high-dimensional survival data using random survival forests (RSF) introduced by Ishwaran et al. in 2008. The study of the behavior of the method is carried out on 4 simulation schemes, the PBC data with and without addition of noise variables and on the actual data on breast cancer. When we optimize the parameters of random forests for the various simulations, we notice that each case is different and that it is therefore important to carry out this step before of the execution of VSURF. For the variable selection, VSURF succeeds in selecting the good variables and few or no noise variables. The VSURF method for high-dimensional survival data is interesting for the variable selection and a method which will be useful, for example, in identifying genes with high prognostic value and predictive power in patients with cancer or other diseases in clinical trials.
Document type :
Master Thesis
Complete list of metadata

https://dumas.ccsd.cnrs.fr/dumas-03377763
Contributor : Isped Ub - Connect in order to contact the contributor
Submitted on : Thursday, October 14, 2021 - 12:02:47 PM
Last modification on : Saturday, October 16, 2021 - 3:22:23 AM

File

M2BIOSTAT_2021_HUCHON.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : dumas-03377763, version 1

Collections

Citation

Mélanie Huchon. Sélection de variables à l’aide de forêts aléatoires pour données de survie de grande dimension. Santé publique et épidémiologie. 2021. ⟨dumas-03377763⟩

Share

Metrics

Record views

26

Files downloads

24