Estimation and optimization of feature selection stability with applications to omics data

Hamer, Victor
(2022)

Files

thesis-victor-hamer-elec.pdf
  • Open Access
  • Adobe PDF
  • 9.45 MB
Download

Details

Authors
  • Hamer, VictorUCLouvain
    author
Supervisors
Dupont, Pierre
Abstract
Feature selection aims to find the most informative input variables for a prediction task. It has become compulsory for a wide variety of applications due to the appearance of high-dimensional data sets. Feature selection is said to be unstable when small changes in data result in large changes in the selected feature sets. The selected variables are often analyzed by domain experts to gain further insights and selection instability strongly limits their sound interpretation. This is a recurrent concern in bio-medical studies where high-dimensional and highly correlated data sets are common and often cause strong instability.  This thesis tackles this instability issue. In particular, selection stability and predictive accuracy are optimized jointly in a bi-objective framework. Pareto-optimal trajectories are derived, from which domain experts can choose a trade-off based on their personal preferences for both objectives.     The thesis also illustrates that current stability measures are unsatisfactory. This is first shown in the context of stability optimization. A novel and more robust stability measure is proposed and incorporates the importance of the selected features in predictive models. Stability estimation is notoriously difficult when the feature space contains strongly correlated features. The thesis also introduces a novel measure in that setting, where computing stability is shown to require solving a constrained optimization problem. Both measures improve decision-making in their respective context. By combining them with new interactive visualization software, this thesis provides domain experts with an enhanced toolbox for domain analysis.
Affiliations

Citations

  • APA
  • Chicago
  • FWB

Hamer, V. (2022). Estimation and optimization of feature selection stability with applications to omics data. https://hdl.handle.net/2078.5/103307