面向流式数据的在线自步多任务特征学习研究

In recent years, A lot of attention has been paid to streaming data analysis from academic community and industry community. Streaming data often have the following characteristics: i) The feature dimension of inputs is often high, and there may be many redundant and noisy features; ii) there are often multiple outputs (or multiple tasks), and there are certain relationship among these outputs (tasks); iii) data is often evolving. In this project, based on these characteristics of streaming data, we will focus on designing a framework of online self-paced multi-task feature learning for dynamic data streams. Specifically, we first build a framework to conduct online multi-task learning and feature learning simultaneously, in order to mine inherent structure of data, and leverage it for improving the model performance. Based on the fact that humans often learn from easy concepts to hard ones in the cognitive process, we will introduce such learning regime to online multi-task learning, making the update of models in a self-paced way. Finally, we will conduct online multi-task feature learning and online self-paced multi-task learning in a unified framework to further improve the performance of the model. This project focuses on designing a series of algorithms for learning the structure of streaming data by leveraging multi-task learning, feature learning, and feature learning, which provides some theoretical insights and practical algorithms for real-world applications.

近年来，流式数据分析得到了学术界和工业界的广泛关注。流式数据常常呈现如下特点：i) 输入数据的特征表示是高维的；ii)输出变量是多维的或者有多个任务同时存在，且变量间或任务间具有某种相关性；iii)数据随着时间的变化会发生"演化"。在本项目中，针对上述流式数据的特点，本申请拟提出在线自步多任务特征学习对流式数据进行挖掘，主要包括：（1）探索建立在线多任务学习和特征学习的统一框架，挖掘流式数据的内在结构；（2）人类在学习的过程中常常先从简单的入手，慢慢再学习复杂的东西。基于这种学习机制，研究在线自步多任务学习算法模型，使得模型学习更加符合人类的学习方式；（3）融合（1）和（2），提出在线自步多任务特征学习的整体框架，进一步提高模型的准确度。本项目立足于借助多任务学习、特征学习和自步学习设计一系列算法，用来挖掘流式数据的内在结构，因此具有理论意义和应用前景。

本项目围绕在线自步多任务特征学习对流式数据进行挖掘，主要包括：（1）建立了在线多任务学习和特征学习的统一框架，挖掘流式数据的内在结构；（2）人类在学习的过程中常常先从简单的入手，慢慢再学习复杂的东西。基于这种学习机制，研究了在线自步多任务学习算法模型，使得模型学习更加符合人类的学习方式；（3）融合（1）和（2），提出了在线自步多任务特征学习的整体框架，进一步提高模型的准确度。共发表学术论文14篇，其中包括中科院JCR-1区期刊或者CCF A类论文10篇。