.png){kind=icon}
.png){kind=icon}
Return
Cross-domain pattern classification with heterogeneous distribution adaptation
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-022-01646-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Heterogeneous domain adaptation (HDA) aims to learn a classification model to label the target samples given the source domain training data, with distinct distribution and difference in feature type or dimension. Most previous HDA methods solve the problem through the mapping of source and target domains into a domain invariant feature subspace to minimize the discrepancies across domains. However, the intrinsic properties of data is lost in such alignment strategy, which reduces the performance of model generalization. In this paper, we propose a novel approch called as Cross-Domain Pattern Classification with heterogeneous distribution adaptation (CDPC). CDPC preserves the intrinsic properties of data by utilizing the sparse coding method to find a new feature representation of features in a shared subspace. Meanwhile, a sample reweighting approach is proposed to align the probability distribution of source and target domain features into a common subspace. Extensive experiments on several HDA tasks including image to image, text to image and text to text illustrate the superiority of our proposed model against other state-of-the-art HDA methods.
Keywords:
Heterogeneous domain adaptation
Sparse subsapce learning
Distribution alignment
Transfer leaning
Journal
IF:
2.7
Papers:
3.1K
Citations:
5.6K