面向机器学习系统的需求建模与决策选择  

Requirements Modeling and Decision-making for Machine Learning Systems

作　　者：杨立[1] 马佳佳 江华禧 马肖肖 梁赓[1] 左春[1,2] YANG Li;MA Jia-jia;JIANG Hua-xi;MA Xiao-xiao;LIANG Geng;ZUO Chun(Institute of Software,Chinese Academy of Sciences,Beijing 100190,China;Sinosoft.Co,Ltd,Beijing 100190,China)

机构地区：[1]中国科学院软件研究所,北京100190 [2]中科软科技股份有限公司,北京100190

出　　处：《计算机科学》2020年第12期42-49,共8页Computer Science

基　　金：中国科学院战略性先导A类专项(XDA20080200);国家重点研发计划项目(2018YFB1005002)。

摘　　要：机器学习支撑的系统应用越来越普遍,但是此类系统的需求通常难以表达完整且可能存在一些难以检测的冲突,使得这些系统通常无法在生产环境中高效满足用户的综合需求。此外,对于在实际场景中使用的机器学习系统,用户信任通常取决于包含可解释性、公平性等非功能需求在内的综合需求的满足程度,且在不同领域内应用机器学习通常有特定的需求,为保证需求描述的质量及实施过程的决策带来了挑战。为解决以上问题,文中提出了一个机器学习系统的需求建模和决策选择框架,包括一个MLS(Machine Learning Systems)需求概念模型和机器学习管道过程元模型,以及对训练数据集、算法等组件的决策选择方法,旨在规范实际场景中机器学习系统的需求设计、开发和评估。实例研究表明,提出的MLS需求描述和实现方法是可行且有效的。The application of systems supported by machine learning is becoming more and more common.However,because the requirements of such systems are often difficult to express completely and there may be some conflicts which are hard to detect,these systems usually cannot efficiently meet the comprehensive needs of users in a real application environment.In addition,for Machine Learning Systems(MLS)used in actual scenarios,user trust usually depends on the satisfaction of comprehensive requirements including non-functional requirements such as interpretability and fairness,and application of machine learning in different fields usually has specific needs,which brings challenges to ensure the quality of requirement description and decision-making for implementation process.To solve above-mentioned problems,this paper presents a machine learning system requirements and decision-making framework which includes a concept MLS requirements model and a Meta-Model of MLS pipeline process,as well as decision making method for training datasets and algorithms selection.The purpose is to standardize the design,development and evaluation of requirements for machine learning used in actual scenarios.The case study shows that the proposed MLS requirement description and implementation method is feasible and effective.

关 键 词：机器学习系统 需求建模 非功能需求 元模型 决策选择 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]