May 2024 – Aug 2024 · 3 mos · Ottawa, Ontario, Canada · On-site

• As a MITACS Globalink Research Intern, I worked under the supervision of Dr. Paula Branco on developing a novel feature selection algorithm tailored for high-dimensional datasets with class complexity, overlap, and imbalance. My research focused on optimizing feature elimination by integrating advanced class complexity and overlap metrics, including F2 and F3 scores, R-values, Augmented R-values, N1, and N3 metrics. These measures allowed for a systematic identification and removal of redundant and noisy features, improving classification performance, model interpretability, and computational efficiency.
• To achieve this, I designed and implemented a voting-based selection mechanism, which combines multiple metrics with adjusted weight distributions to optimize feature elimination. This approach was rigorously tested on 30 publicly available datasets and 86 synthetic datasets, each containing informative, non-informative, and noisy features. The results demonstrated superior performance compared to conventional feature selection methods, including Chi-Squared, SVM-RFE, and RELIEF-F.
• This research contributes to machine learning and data science by providing a robust preprocessing step for improving classification models, particularly in domains where class distributions overlap, feature redundancy is high, and imbalance is present.