众包环境下的成对约束传递问题研究

项目名称： 众包环境下的成对约束传递问题研究

项目编号： No.61300164

项目类型： 青年科学基金项目

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： 付振勇

作者单位： 南京邮电大学

项目金额： 25万元

中文摘要： 成对约束作为一类弱监督信息已经被广泛用于各类机器学习、模式识别问题中，成对约束传递可以有效地利用有限的约束信息，但现有的传递算法并没有很好地解决初始约束的获取问题，如何得到廉价的初始约束，如何能够在初始约束包含大量噪声的情况下实现约束信息的有效传递都是值得探索的重要问题。本项目在申请人前期关于成对约束传递问题研究的基础上，充分利用新型的大众协作方式众包获取大量廉价的初始约束信息。众包环境下获得的初始约束具有多源、含大量噪声的特点。我们将利用约束张量来统一表示多源初始约束，在低秩张量恢复和多维信息扩散框架下处理多源约束信息、消除噪声影响并利用大量未约束数据实现众包约束信息的传递。该项目的研究充分考虑了复杂的众包环境下获取的成对约束的特点，提出了针对性的方案，将会解决现有约束传递算法对获取初始约束的限制，极大地扩展约束传递算法的应用，同时也为众包分析的研究提供有益的补充。

中文关键词： 成对约束；约束传递；低秩分析；图像分类；图像聚类

英文摘要： As a weaker type of supervisory information, pairwise constraints have been widely used for many machine learning and pattern recognition problems. Pairwise constraint propagation can effectively exploit the scarce pairwise constraints. However, the current propagation methods cannot address the problem of how to obtain the initial constraints. It is worth to explore how to obtain the cheap initial constraints and how to effectively propagate the constraints when there exists a lot of noise in the initial constraints. On the basis of our previous research works about pairwise constraint propagation, this project will utilize a new kind of crowd cooperation manner, crowdsourcing, to obtain plenty of cheap initial constraints. The initial constrains obtained in the crowdsourcing environment are from multiple providers and include a lot of noise. We will apply the constraint tensor to express the initial constraint from multiple providers. Under the framework of low-rank tensor recovery and multidimensional information spreading, we will handle the constraint from multiple providers, remove the noise and exploit the large amount of unconstrained data to implement the propagation of the initial constraints from crowdsourcing. The research of this project takes into full account the characteristic of the constraints

英文关键词： pairwise constraint；constraint propagation；low-rank analysis；image classification；image clustering