[1]翟利志,任一夫,白 洁,等.基于传递式领域自适应的异构样本增强方法[J].计算机技术与发展,2024,34(01):17-22.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 003]
　ZHAI Li-zhi,REN Yi-fu,BAI Jie,et al.Heterogeneous Sample Enhancement Based on Transitive Domain Adaptation[J].,2024,34(01):17-22.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 003]

点击复制

基于传递式领域自适应的异构样本增强方法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年01期

页码:

17-22

栏目:

大数据与云计算

出版日期:

2024-01-10

文章信息/Info

Title:

Heterogeneous Sample Enhancement Based on Transitive Domain Adaptation

文章编号:

1673-629X(2024)01-0017-06

作者:

翟利志¹; 2 ; 任一夫¹; 2 ; 白 洁¹ ; 高学攀¹ ; 贾庆超¹ ; 刘 强³

1. 中国电子科技集团公司第五十四研究所,河北 石家庄 050081;

2. 河北省智能化信息感知与处理重点实验室,河北 石家庄 050081;

3. 陆装驻石家庄地区第一军代室,河北 石家庄 050081

Author(s):

ZHAI Li-zhi1; 2 ; REN Yi-fu1; 2 ; BAI Jie1 ; GAO Xue-pan1 ; JIA Qing-chao1 ; LIU Qiang3

1. The 54th Research Institute of CETC,Shijiazhuang 050081,China;

2. Hebei Key Laboratory of Intelligent Information Perception and Processing,Shijiazhuang 050081,China;

3. The First Military Office in Shijiazhuang,Shijiazhuang 050081,China

关键词:

域适应; 样本增强; 迁移学习; 小样本; 数据驱动建模

Keywords:

domain adaptation; sample enhancement; transfer learning; small sample; data-driven modeling

分类号:

TP39

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 01. 003

摘要:

小样本问题广泛存在于数据驱动建模。 领域自适应方法通过将源域中的样本知识迁移到目标域,从而实现目标域中的小样本增强,然而此类方法在实际应用中受限,原因在于难以应对领

域分布差异较大的样本增强场景。 针对上述问题,该文提出基于传递式领域自适应的异构样本增强方法。 首先,提出传递式探索策略,通过私有特征和共享特征设计了面向异构域的领域分布探索策略,有效地缓解了负迁移,并为后续分布匹配提供支撑;然后,提出分布联合匹配机制,通过联合匹配异构领域的边缘分布和条件分布,并嵌入自适应机制,从而保证了异构域分布的匹配精度。 该方法在业界公认的田纳西-伊斯曼数据集进行验证,实验结果表明该方法在异构域中的建模表现优于其他方法。

Abstract:

Small sample problem exists widely in data-driven modeling. Domain adaptation achieves small sample enhancement in targetdomain by transferring sample knowledge from source domain to target domain. However, those methods are limited in practicalapplication because it is difficult to deal with sample enhancement scenarios with large domain distribution differences. To solve theseproblems,we propose a heterogeneous sample enhancement method based on transitive domain adaptation. Firstly,a transitive explorationstrategy is proposed. A domain distribution exploration strategy for heterogeneous domains is designed based on specific and commonfeatures,which effectively alleviates negative transfer and provides support for subsequent distribution matching. Then,a distributed jointmatching mechanism is proposed to match the marginal distribution and conditional distribution of heterogeneous domain,and embed anadaptive mechanism to ensure the matching accuracy of heterogeneous domain distribution. The proposed method is verified by theindustry-recognized Tennessee-Eastman dataset,and the experimental results show that the proposed method performs better than othermethods in heterogeneous domain modeling.

常用功能