面向多媒体排序学习的维数约简

项目名称： 面向多媒体排序学习的维数约简

项目编号： No.61271325

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 无线电电子学、电信技术

项目作者： 冀中

作者单位： 天津大学

项目金额： 70万元

中文摘要： 由于多媒体数据特征维数很高，使得排序学习与其他学习算法一样受维数灾难等问题的影响，致使算法的泛化能力和效率严重受限。为了有效克服排序学习中的这些问题，现有方法通常直接使用常规的维数约简策略。但其输入仅是高维数据及类别标签，不能充分利用训练数据的等级、顺序等重要的排序信息作为算法的输入，因而不能达到最优的维数约简效果。为了解决这一严重问题，本项目提出了将排序信息作为输入并充分利用排序学习机制的新型维数约简技术（简称"排序维数约简"）。首先重点研究有效利用排序信息替代类别标签进行维数约简的方法；接着进一步结合排序学习的独特学习机制研究相应维数约简方法；然后作为拓展研究，探索与半监督学习、迁移学习等方法相结合的途径与方法；最后结合多媒体检索、个性化推荐等实际应用验证算法的先进性和实用性。本项目的实施将丰富和完善排序学习和维数约简的基础理论研究，推动多媒体搜索、移动互联网等相关产业的发展。

中文关键词： 维数约简；排序学习；特征提取；多媒体信息检索；排序信息

英文摘要： Due to the high feature dimensionality of multimedia data, learning to rank approaches are prone to problems such as "curse of dimensionality", which is a common phenomenon occurred in machine learning algorithms and will deteriorate the generalization ability and the efficiency seriously. To efficiently overcome these problems in learning to rank system, current methods generally adopt the conventional dimensionality reduction strategy. However, the inputs of this kind of strategy are only the high-dimensional data and their labels, leaving the ranking information such as the relevance score and the priority aside. Therefore, the current dimensionality reduction strategy cannot achieve the best performance. To solve this kind of critical problem, the research project proposes a novel technology named as "Ranking Dimensionality Reduction", which can not only take the ranking information as input, but also make full use of the unique learning mechnism in learning to rank methods. Specifically, the research first lays emphasis on the effective methods of replacing label information with ranking information as the input of dimensionality redcution approaches. Further, the research inverstigates the effective utilization of ranking information together with the unique rank learning mechnism. And then, as an extended

英文关键词： Dimensionality Reduction；Learning to Rank；Feature Extraction；Multimedia Information Retrieval；Ranking Information