Unified Cross-domain Classification via Geometric and Statistical Adaptations
作者：
Highlights：
• To deal with the distribution divergence between domains, we propose a domain adaptation model UCGS based on the coupled adaptations theory. UCGS combines the inter-domain distribution divergence reduction and classifier construction in a unified model for robust transfer learning.
• UCGS employs MMD to formalize the distribution divergence statistically. The means of the data distributions are well matched through minimizing MMD.
• Furthermore, UCGS flexibly employs the Nyström method to explore the inter-domain geometric connections and uses the Nyström approximation error to quantify the inter-domain geometric differences. A domain-invariant graph is finally constructed to bridge two domains geometrically.
• Comprehensive experiments on real-world datasets verify the superiority of UCGS.
摘要
•To deal with the distribution divergence between domains, we propose a domain adaptation model UCGS based on the coupled adaptations theory. UCGS combines the inter-domain distribution divergence reduction and classifier construction in a unified model for robust transfer learning.•UCGS employs MMD to formalize the distribution divergence statistically. The means of the data distributions are well matched through minimizing MMD.•Furthermore, UCGS flexibly employs the Nyström method to explore the inter-domain geometric connections and uses the Nyström approximation error to quantify the inter-domain geometric differences. A domain-invariant graph is finally constructed to bridge two domains geometrically.•Comprehensive experiments on real-world datasets verify the superiority of UCGS.
论文关键词：Domain adaptation,Statistical adaptation,Maximum mean discrepancy (MMD),Geometric adaptation,Nyström method
论文评审过程：Received 18 May 2020, Revised 27 July 2020, Accepted 9 September 2020, Available online 10 September 2020, Version of Record 1 November 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107658
