Joint discriminative subspace and distribution adaptation for unsupervised domain adaptation

摘要

In traditional machine learning algorithms, the classification models are learned on the training data (source domain) to reuse for labelling the test data (target domain) where the training and test samples are from the same distributions. However in nowadays applications, the existence of distribution shift across the source and target doamins degrades the model performance, significantly. Domain adaptation methods have been proposed to compensate domain shift problem by aligning the distributions across the source and target domains under various adaptation strategies. This paper addresses the robust image classification problem for unsupervised domain adaptation. Specifically, following three methods are proposed: Discriminative Subspace Learning (DSL), Joint Geometrical and Statistical Distribution Adaptation (GSDA), and Joint Subspace and Distribution Adaptation (DSL-GSDA). DSL is a subspace centric method that aligns the specific and shared features across domains. Indeed, DSL finds two projections to map the source and target data into independent subspaces by aligning the discriminant and global structures of domains. GSDA trends to find an adaptive classifier through statistical and geometrical distribution alignment and minimizes the prediction error. DSL-GSDA, as a combination of DSL and GSDA, consists of two subspace and distribution adaptation levels. DSL-GSDA uses DSL to build two aligned subspaces of source and target domains. The distributions of source and target data in new subspaces is adapted via GSDA. The proposed methods are evaluated on benchmark visual datasets for object, digit and face recongnition tasks. Visual datasets consist of image domains that have been captured under various real-world conditions where the domain shift is unavoidable. The experiment results show that DSL, GSDA and DSL-GSDA outperform other state-of-the-art domain adaptation methods by 6.19%, 1.48% and 1.99% improvement, respectively. Our source code is available at https://github.com/jtahmores/DSLGSDA (https://github.com/jtahmores/DSLGSDA).