Locality Fisher discriminant analysis for conditional domain adaption

Abstract

Domain adaptation tackles a learning problem in a test data (target domain) by utilizing the training data (source domain) in a related domain, often with different distribution. Intuitively, discovering a good feature representation across the source and target domains is crucial to boost the model performance. In this paper, we are to find such a representation through a novel learning method, locality Fisher discriminant analysis for conditional domain adaption (LFDA-CDA). LFDA-CDA tries to find a linear combination of features across domains in an embedded subspace using Bregman divergence (BD). In the embedded subspace spanned with new representation, data distributions in different domains are close to each other. Therefore, the standard machine learning methods can be applied to train a model on the source data for exploiting on the target data. In fact, we propose a novel feature representation in which to perform conditional domain adaptation via a new parametric nonlinear BD method, which can considerably minimize the distributions mismatch across domains by projecting data onto the learned subspaces. Moreover, LFDA-CDA benefits from data geometric structure segmentation and alignment between source and target domains, through locality preserving projection. Extensive experiments on 16 real vision datasets with different difficulties verify that LFDA-CDA can significantly outperform state-of-the-art methods in image classification tasks. Our source code is available at https://github.com/Jtahmores/LFA-CDA .