项目名称: 基于社会标记精化的多标记学习算法研究
项目编号: No.61202170
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 卫志华
作者单位: 同济大学
项目金额: 24万元
中文摘要: 不同于传统的单标记学习问题,多标记学习考虑一个对象对应多个类别标记的情况,是当前国际机器学习领域研究的热点问题之一。然而,由于多标记学习中标记组合的多样性,其学习的准确率尚待提高,并且训练样本获取困难问题已成为制约该领域研究的瓶颈。本项目以互联网视频自动标注为背景,研究以下问题:1)针对社会标记的模糊性、个性化等问题,拟通过内容感知与语义计算相结合的方法,对社会标记进行去噪、修正、推荐、合并,实现社会标记精化;2)针对标记关系不确定性的问题,拟通过贝叶斯网络对标记关系建模,实现标记关系的量化描述;3)针对多标记学习中常忽略标记依赖关系的问题,拟通过最优化理论训练标记关系对多标记集成学习的约束参数,达到提高多标记分类性能的目的。不同于传统的直推式学习方法,采取将样本精化和多标记学习两个复杂问题分而治之的思路,丰富和拓展了机器学习理论和方法,对于解决互联网视频标注问题具有重要的指导意义。
中文关键词: 多标记学习;社会标记精化;视频内容标注;半监督学习;目标跟踪
英文摘要: Multi-label learning, which considers the case of an object related to multiple labels, attracts much attention in recent years. However, the performance of multi-label learning is still not satisfied due to its own complexity. Moreover, the scarcity of training samples for learning process becomes the necklace of the research for the great price of manually labeling works. Focusing on the automatic video content annotation, this project studies the following problems. Firstly, considering the problem that the social tags contributed by common users on the Web are often ambiguous, limited in terms of completeness,and overly personalized, the social tag refining algorithm is proposed based on content perception and semantic computing.Secondly, considering the uncertainty relationship among labels, the algorithm based on Bayesian Networks is proposed to modeland describe their relationship.In addition, considering the effect of label correlation, the algorithm of ensembling multiple single classifiers constrained by label dependency is proposed.Different from the traditional transductive learning methodology, this proposal adopts the strategy of "divide and rule" which resolve the problem of refining social tags and advancing multi-label learning performance respectively. It enriches and extends the theories and a
英文关键词: Multi-label learning;society label refining;video content annotation;semi-supervised learning;object tracking