Combinatorial Methods for Feature Selection and Biomarker Discovery

Project Description: This research explores feature selection through combinatorial optimization techniques, particularly using domination and packing in graphs and networks. By modeling feature selection as a graph-theoretic problem, we use structures such as independent dominating sets to identify minimal representative features and packing formulations to promote diversity while reducing redundancy. Our approach aims to identify compact representative feature sets while reducing redundancy and preserving predictive power. We also investigate parameterized complexity aspects and specialized heuristics for large-scale data. Applications include bioinformatics, biomarker discovery, disease prediction, and high-dimensional data classification. Recent work investigates the identification of compact biomarker signatures from high-dimensional biomedical data. Preliminary experiments on public cancer datasets, including leukemia and colon cancer datasets, suggest that substantial feature reduction can be achieved while maintaining strong predictive performance.

PI: Faisal N. Abu-Khzam

Partners: Joseph Barr and Peter Shaw.

Publications:

F. N. Abu-Khzam, J. R. Barr, M. R. Benabid and P. Shaw. Feature Selection via Weighted Independent Domination, in proceedings of the 2024 Conference on AI, Science, Engineering, and Technology (AIxSET): 2024, 179-184.

Recent Manuscript:

Pack and Extract: An Effective Graph-Based Approach for Feature Selection

This work investigates the application of graph-based feature selection to biomarker discovery and disease prediction from high-dimensional biomedical data.