面向异构环境的多任务多视图学习算法研究

项目名称： 面向异构环境的多任务多视图学习算法研究

项目编号： No.61473273

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 庄福振

作者单位： 中国科学院计算技术研究所

项目金额： 78万元

中文摘要： 本项目针对异构环境下的数据特点，深入分析多任务学习的挑战性问题，系统地对多任务多视图学习算法进行研究。通过考察不同任务之间的相关性，即多个任务可能只是部分相关而不是所有任务都相关，探讨基于深度学习的任务间相关性关系度量,并研究基于狄利克雷过程的聚类多任务学习算法避免不适当的知识共享。异构环境下的多个任务数据通常呈现多种多样性，即它们包含的样本类别空间和特征空间可能都不一致。研究基于主题模型和判别分析的多任务多视图学习算法解决任务间含有不同样本类别空间的学习问题；研究提出基于图模型的多任务多视图学习算法解决多个任务具有不同特征空间的学习问题。为了满足海量数据的处理需求以及实际应用，研究基于Spark 的高效并行多任务多视图学习算法。预期在IEEE TKDE、IEEE TOC 等重要国际期刊，以及SIG KDD、IJCAI、AAAI、ACM CIKM等重要学术会议上发表论文20篇。

中文关键词： 多任务学习；异构环境；共享结构；半监督学习；多视图学习

英文摘要： In this project, we will deeply analyze the challenging problems of multi-task learning based on the data characteristics in heterogeneous environment, and systematically study the multi-task multi-view learning algorithms. Through investigating the relatedness of multiple tasks, they may be related to each other partly, rather than that all tasks are globally related to each other. Thus, we will investigate the relatedness measure among multiple tasks via deep learning techniques, and study the clustered multi-task learning algorithms based on Dirichlet process to avoid inappropriate knowledge sharing. Due to the diversity of multi-task data characteristics in heterogeneous environment, the sample category space and feature space among multiple tasks may not be consistent. To address these challenges, we will systematical study the multi-task multi-view learning algorithms applying the techniques of topic models, discriminant analysis and graph models etc. To meet the requirement of real-world applications, we will also design and implement parallel multi-task multi-view learning algorithms based on the Spark coding framework to process large-scale data sets. We expect to publish twenty papers in the important international journals, such as IEEE TKDE, IEEE TOC and so on, or important international conferences, such as SIG KDD, IJCAI, AAAI, ACM CIKM and so on.

英文关键词： Multi-task Learning;Heterogeneous Environment;Shared Structure;Semi-supervised Learning;Multi-view Learning