Cross-domain structure preserving projection for heterogeneous domain adaptation

Heterogeneous Domain Adaptation (HDA) addresses the transfer learning problems where data from source and target domains are of different modalities (e.g., texts images) or feature dimensions features extracted with methods). It is useful for multi-modal analysis. Traditional domain adaptation algorithms assume that representations samples reside in same space, hence likely to fail solving heterogeneous problem. Contemporary state-of-the-art HDA approaches usually composed complex optimization objectives favourable performance therefore computationally expensive less generalizable. To address these issues, we propose a novel Cross-Domain Structure Preserving Projection (CDSPP) algorithm HDA. As an extension classic LPP domains, CDSPP aims learn domain-specific projections map sample into common subspace such class consistency preserved distributions sufficiently aligned. simple has deterministic solutions by generalized eigenvalue naturally suitable supervised but also been extended semi-supervised unlabelled available. Extensive experiments have conducted on commonly used benchmark datasets (i.e. Office-Caltech, Multilingual Reuters Collection, NUS-WIDE-ImageNet) as well Office-Home dataset firstly introduced ourselves due its significantly larger number classes than existing ones (65 vs 10, 6 8). The experimental results both demonstrate superior our proposed method against contemporary methods.