开放动态环境下在线机器学习理论与方法

项目名称： 开放动态环境下在线机器学习理论与方法

项目编号： No.61333014

项目类型： 重点项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 周志华

作者单位： 南京大学

项目金额： 290万元

中文摘要： 传统的机器学习研究主要是在封闭静态环境下展开的，研究者们通常假设训练样本分布与测试样本分布相同、样本类别和属性恒定、学习目标明确唯一。随着机器学习技术逐渐走向实用，不可避免地在越来越多的任务中面临开放动态环境；在这样的环境中，训练/测试样本分布可能不同、样本类别可能增加、属性可能增加/失效、学得模型要尽可能满足不同用户的需求。本项目针对开放动态环境“分布偏移”、“类别可增”、“属性变动”和“目标多样”这四个关键因素，并由于数据量大且不断积聚而需要采用在线学习方式，拟从理论上分析上述因素对可学习性的影响，提出能有效适应分布偏移的在线学习方法、能有效适应类别增加的在线学习方法、能有效应对属性变动的在线学习方法、以及能有效适应多样化目标的在线学习方法。基于上述研究工作，本项目将在国内外重要期刊和会议发表论文15-25 篇，申请发明专利3-5 项，研制原型系统1个，培养多名博士后、研究生。

中文关键词： 机器学习;开放动态环境;在线学习;泛化;学习理论

英文摘要： Conventional machine learning researches are generally conducted in closed and static environments. These researches usually assume that the training data distribution and test data distribution are identical, the data categories and attributes are fixed, and the learning objective is clear and unique. As machine learning techniques come to real practice, open and dynamic environments are encountered in more and more tasks inevitably. In such environments, the training/test distributions can be different, the number of categories can increase, the attributes can augment or fail, the learned model need to meet different requirements of users. In this project, we will focus on the four key factors, that is, “drifting distributions”, “augmentable categories”, “varying attributes” and “various objectives”, and consider online learning style by noticing the large and accumulating data in open and dynamic environments. We plan to theoretically analyze the influences of these factors on learnability, and propose online learning approaches that are able to adapt to drifting distributions, augmentable categories, varying attributes and various objectives. As the result of this project, 15-25 high quality papers are expected to be published in leading journals and conferences, as well as 3-5 patents and a prototype system, and a number of postdocs and graduate students will be trained.

英文关键词： Machine Learning;Open Dynamic Environment;Online Learning;Generalization;Learning Theory