光華講壇——社會名流與企業家論壇第6590期

主題：A unified performance measurement framework for classification algorithms 一種針對分類算法的統一性能度量框架

主講人：美國特拉華大學 陳濱桐教授

主持人：工商管理學院 李偉教授

時間：10月11日 14:30-16:30

舉辦地點：頤德樓H308

主辦單位：工商管理學院 科研處

主講人簡介：

Bintong Chen graduated from Shanghai Jiaotong University with dual B.S. degrees in ship-building/naval architecture and electrical engineering. He received M.S. in systems engineering and Ph.D. in operations management/research from the Wharton School, the University of Pennsylvania. He is currently a professor of the Lerner College of Business and Economics and the director of the Institute for Financial Services Analytics at University of Delaware. He published many high quality papers in the area of optimization theory, data-driven analytics, and business applications. He received many outstanding research and teaching awards in institutions he worked. Professor Chen consulted many international companies, including JP Morgan Chase, Agriculture Bank of China, AT&T, Burlington Northern Rail, Delaware Department of Transportation, Nordstrom, and AstraZeneca, etc. He was a board member for APICS, the largest supply chain professional association in North American.

陳濱桐，特拉華大學勒納商學與經濟學院教授，于1985年獲上海交通大學船舶結構與海洋工程系、電子工程系雙學士學位，1987年獲賓夕法尼亞大學工程與應用科學學院系統工程系碩士學位，1990年獲賓夕法尼亞大學沃頓商學院運籌與信息管理系博士學位。現為特拉華大學勒納商學與經濟學院金融服務分析中心主任、博士項目主任。陳濱桐教授在管理科學、運籌學和運營管理領域取得了豐碩的研究成果，有多項研究成果發表在相關領域全球頂級學術期刊，包括《Management Science》和《Operations Research》，曾在諸多國際期刊編輯委員會中任職，包括《POM》和《Omega》期刊，并在所工作的大學中獲得了許多杰出的研究和教學獎項。陳教授曾為許多國際公司提供咨詢服務，包括摩根大通、中國農業銀行、美國電話電報公司、伯靈頓北方鐵路、特拉華州交通部、諾德斯特龍和阿斯利康等。他曾是北美最大的供應鏈專業協會——美國生產與庫存管理協會(APICS)的董事會成員。

內容簡介:

Many classification algorithm performance measures have been independently proposed and studied. Two questions arise about these measurements: (1) When do they measure the maximum potential of a classification algorithm? (2) How to efficiently identify and calculate the maximum performance for each measurement? We propose a unified theoretical framework that includes all existing performance measures and curves as special cases. To answer the first question, we investigate two variable transformations and apply theoretical findings to various measures and performance curves. To answer the second question, we classify all performance measures into three categories: monotone measures, unimodal measures, and multi-modal measures, based on the process to search for the optimal threshold. The unified framework allows us to systematically analyze the properties of classification algorithm performance measures and provides guidance to design new performance measures.

許多分類算法的性能度量方法都是被獨立提出和研究。關于這些度量方法存在兩個問題：（1）它們在何時度量分類算法的最大潛力？（2）如何高效識別并計算每個度量方法的最大性能？我們提出了一個統一的理論框架，其中包括所有現有的性能度量方法和曲線作為特例。為回答第一個問題，我們研究了兩個變量變換，并將理論發現應用到了各種度量方法和性能曲線上。為回答第二個問題，我們基于搜索最優閾值的過程將所有性能度量方法分為了三類：單調度量、單峰度量和多峰度量。這個統一的框架使我們能夠系統地分析分類算法性能度量的屬性，并為設計新的性能度量方法提供指導。