Public Thesis defense - ICTEAM

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

Pour l’obtention du grade académique de Doctorat en sciences de l’ingénieur et technologie

Feature selection aims at finding the most informative input variables for a prediction task such as classification or regression. Current selection methods, especially applied to high dimensional data, tend to suffer from instability since marginal modifications in the data may result in largely distinct selected feature sets. This instability strongly limits a sound interpretation of the selected variables by domain experts. The lack of trust from domain experts certainly hinders the expansion of machine learning in sensitive areas such as the biomedical domain.

In this thesis, I address this issue by developing tools to properly estimate and optimize selection stability. In particular, selection stability is optimized along with predictive accuracy in a bi-objective framework. Pareto-optimal trajectories are derived, from which domain experts can choose a trade-off based on their personal preferences towards both objectives. The thesis further shows that current stability measures suffer from limitations. More robust alternatives are discussed, both in the context of stability optimization and when the feature space contains highly correlated variables, which poses specific challenges.

The different tools proposed in the thesis improve the quality of the domain analysis and the expert’s trust in machine learning, which should ease its adoption in sensitive domains.

Jury members :

• Prof. Pierre Dupont (UCLouvain), supervisor
• Prof. Michel Verleysen (UCLouvain), chairperson
• Prof. Charles Pecheur (UCLouvain), secretary
• Prof. Benjamin Haibe-Kains (University of Toronto, Canada)
• Prof. Yvan Saeys (UGent)
• Dr. Cyril de Bodt (UCLouvain)

Pay attention :

The public defense of Victor Hamer scheduled for Wednesday 07 December at 03:00 p.m. will also take place in the form of a video conference

Télécharger l'annonce