基于多任务学习的多源数据分类研究  

Research of classification method for multi-source data based on multi-task learning

作　　者：马建阳 张宝鹏[1] Ma Jianyang;Zhang Baopeng(School of Computer&Information Technology,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044

出　　处：《计算机应用研究》2018年第11期3228-3231,共4页Application Research of Computers

基　　金：中央高校基本科研业务费资助项目(2017JBM320);国家科技支撑计划课题资助项目(2014BAH24F02)

摘　　要：针对现有方法在处理多源数据时忽视数据源之间关联性的问题,提出了一种可以同时实现多分类效果的多源学习框架。该框架将不同的数据源看做多个相关的任务,将多源问题转换为经典的多任务学习问题,通过提取数据源之间的关联,来提高单个数据源的分类性能。此外,该框架利用聚类分析原理,对带标记样本实现多分类效果。实验结果表明,该框架优于只针对单个数据源学习的单任务学习框架和只针对二分类进行处理的传统的多任务学习框架。Since the existing models often suffer from the problem of ignoring the relationship between data sources,this paper proposed a multi-source learning framework which could simultaneously achieve multi-class classification effects.By taking different data sources as multiple related tasks,it transformed multi-source problems into classic multi-task learning problems.Thus,this framework enhanced the classification performance of single data source by extracting relationship between data sources.In addition,the framework utilized clustering analysis for the tagged samples to achieve multi-class classification.The experimental results show that the proposed framework can increase the classification accuracy than that of single-task learning framework aiming at single data source and traditional multi-task learning framework aiming for binary classification problems.

关 键 词：多源学习 多分类 任务相关性 多任务学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]