國立政治大學統計學系
學 術 演 講
學 術 演 講
主講人：魏裕中副教授(國立彰化師範大學統計資訊研究所)
題 目：A Logic Tree Ensemble Balancing Interpretability and
Predictive Performance for Feature Interaction Discovery
in High-Dimensional Data
時 間：民國114年11月3日 (星期一) 下午1：30
地 點：國立政治大學逸仙樓050101教室
摘 要：
Identifying meaningful feature interactions in high-dimensional data remains a central challenge in modern statistical learning. Traditional statistical models are interpretable but often require pre-specified interaction forms, limiting their ability to capture complex patterns in large feature spaces. In contrast, machine learning models offer strong predictive performance and flexibility in uncovering intricate interactions, yet their limited transparency can hinder adoption in domains where interpretability is essential.
Logic regression, a statistical method based on logic trees that represent Boolean combinations of features, provides an interpretable framework for modeling interactions. However, in high-dimensional contexts such as genome-wide association studies, where only a small subset of features is likely to be informative, logic regression may struggle to identify relevant interactions, leading to reduced predictive accuracy.
To address these limitations, we propose the iterative weighted Logic Forest (iwLF), a hybrid ensemble algorithm that integrates principles from both statistics and machine learning. By employing logic trees as base learners and introducing two distinct weighting schemes, iwLF enhances both interpretability and predictive capability.
Empirical results on both simulated and real-world datasets demonstrate that iwLF tends to achieve more stable prediction performance and improved identification of important features and their interactions, compared to other logic tree-based approaches.
