A robust and interpretable feature selection pipeline

• Main Author: Krishnan, Nithya
• Other Authors: A S Madhukumar
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/148109
• Institution: Nanyang Technological University

A feature selection pipeline that removes redundant and irrelevant features without resulting in a significant drop in performance is investigated in this work. The novel pipeline frameworks consider the combined effect of redundancy minimisation through Principal Feature Analysis (PFA) algorithms and relevant feature selection through Causality-Based(Causality-based) methods. These independent methods and pipeline frameworks undergo a comprehensive evaluation upon diverse datasets using a variety of evaluation metrics. It is demonstrated that such methods can significantly decrease the number of features while maintaining a less than proportional drop in performance. The pipelines are also built to be interpretable, with the user being able to know which features are removed at each stage of the pipeline and the reasons for doing so. Pipeline frameworks which incorporate Causality-based methods followed by PFA methods are also computationally efficient and do not take a considerable amount of time. These frameworks also improve upon the performance of the independent PFA and Causality-based methods used, providing a promising tool for interpretable and robust feature selection.