Content area
Full Text
Received 23 February 2015; received in revised form 16 April 2016; accepted 4 July 2016
KEYWORDS
Multi-source transfer learning; Domain adaptation; Discriminative dimensionality reduction; Fisher discriminant analysis.
Abstract. Transfer learning is a well-known solution to the problem of domain shift in which source domain (training set) and target domain (test set) are drawn from different distributions. In the absence of domain shift, discriminative dimensionality reduction approaches could classify target data with acceptable accuracy. However, distribution difference across source and target domains degrades the performance of dimensionality reduction methods. In this paper, we propose a Discriminative Dimensionality Reduction approach for multi-source Transfer learning, DiReT, in which discrimination is exploited on transferred data. DiReT finds an embedded space, such that the distribution difference of the source and target domains is minimized. Moreover, DiReT employs multiple source domains and semi-supervised target domain to transfer knowledge from multiple resources, and it also bridges across source and target domains to find common knowledge in an embedded space. Empirical evidence of real and artificial datasets indicates that DiReT manages to improve substantially over dimensionality reduction approaches.
© 2017 Sharif University of Technology. All rights reserved.
(ProQuest: ... denotes formulae omitted.)
1. Introduction
In machine learning and pattern recognition, dimensionality reduction is the process of reducing the number of features via obtaining a collection of principle variables [1-3]. Fisher Discriminant Analysis (FDA) [4] and Principal Components Analysis (PCA) [5] are pioneer approaches that concentrate on discovering a low-dimensional latent space. However, there are many real-world applications whose conditions for developing and using the models are different. In this case, the embedding for source and target domains is drawn from different distributions; therefore, the performance of model degrades dramatically.
Domain shift or data shift is a common challenge in real-world applications in which training and test sets have different distributions. This problem arises of a variety of applications, such as computer vision [6-9], multivariate time series [10], and sentiment analysis [11,12].
In this paper, an invariant latent space is extracted to tackle domain shift problem. DiReT, Discriminative Dimensionality Reduction approach for multi-source Transfer learning discovers a latent space, which is discriminative between different classes. DiReT employs Fisher discriminant analysis to find domain invariant features across source and target domains in a semi-supervised...