國立政治大學統計學系
學 術 演 講
主講人：Jin-Chuan Duan (段錦泉院士)學 術 演 講
Chairman of Criat and ADBIZA
Professor Emeritus, National University of Singapore
國立政治大學國際金融學院兼任講座教授
題 目：Machine Learning Interpretable Financial Models via Sequential Monte
Carlo Optimization
時 間：民國113年6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：
Utilizing big data without a theoretical or commonly accepted conceptual basis has been proven technically feasible and may work well in many applications. But black-box machine learning typically fails to meet managerial needs and/or compliance requirements due to its lack of interpretability in a conventional sense. In principle, one could directly machine-learn interpretable conventional models instead of searching for ways to interpret black boxes afterwards. However, Standard approaches to constructing conventional models are ill-equipped to handle high-dimensional data arising from, say, digital footprints. This talk is about expanding the realm of interpretable models through machine learning via sequential Monte Carlo (SMC) optimization. First, I will illustrate the general idea behind SMC optimization and show its working using a simple example. Next, I will demonstrate how to construct a conventional hedonic housing price model through finding an optimal stable subset of interpretable features out of over 190,000 potential variables arising from interacting standard data features. The resulting parsimonious model is not only naturally interpretable but also outperforms large neural networks. Finally, I will discuss my ongoing research work on finding the optimal stable decision tree in a random forest.
Background reading: "Sequential Monte Carlo optimization and statistical inference" Duan, J.-C., Li, S., & Xu, Y. (2022). Wiley Integrative Reviews: Computational Statistics, e1598. https://doi.org/10.1002/wics.1598.
