Linear Discriminative Projections for Heterogeneous Domain Adaptation
Date Issued
2014
Date
2014
Author(s)
Fang, Wen-Chieh
Abstract
It is often expensive to collect labeled data and we sometimes have large
amounts of labeled data in a related domain. Without enough training data,
some classifiers such as k-Nearest Neighbor (kNN) or Support Vector Machine
(SVM) may fail to achieve good classification performance. In this
thesis, we consider the problem of utilizing few labeled data samples in a
target domain and the data samples in a source domain to improve data classification
in the target domain. We assume that the source and target domains
have different feature spaces. In addition, the two domains are assumed to
share no explicit common features but have the same set of class labels.
A key technique for leveraging the data from another domain is to find
two mapping functions so that the source and target spaces can be projected
on a common space. In this thesis, we present a simple and intuitive technique
called linear discriminative projections to address the problem. First, we separate
the source data of distinct classes by using a discriminative method such
as Linear Discriminative Analysis (LDA). We then apply a regression technique
to map each labeled target data instance as close as possible to the center
of the source data group with the same class label. Finally, we again use a
discriminative method to separate all the data of distinct classes. Experimental
results on some benchmark datasets clearly demonstrate that our approach
is effective for learning discriminative features for supervised classification
with few training target data.
amounts of labeled data in a related domain. Without enough training data,
some classifiers such as k-Nearest Neighbor (kNN) or Support Vector Machine
(SVM) may fail to achieve good classification performance. In this
thesis, we consider the problem of utilizing few labeled data samples in a
target domain and the data samples in a source domain to improve data classification
in the target domain. We assume that the source and target domains
have different feature spaces. In addition, the two domains are assumed to
share no explicit common features but have the same set of class labels.
A key technique for leveraging the data from another domain is to find
two mapping functions so that the source and target spaces can be projected
on a common space. In this thesis, we present a simple and intuitive technique
called linear discriminative projections to address the problem. First, we separate
the source data of distinct classes by using a discriminative method such
as Linear Discriminative Analysis (LDA). We then apply a regression technique
to map each labeled target data instance as close as possible to the center
of the source data group with the same class label. Finally, we again use a
discriminative method to separate all the data of distinct classes. Experimental
results on some benchmark datasets clearly demonstrate that our approach
is effective for learning discriminative features for supervised classification
with few training target data.
Subjects
資料映射
特徵學習
領域適應
監督式分類
機器學習
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-103-D97922009-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):c08196be13ec1f0be95ca257b3c3a63c