Heterogeneous Tensor Domain Adaptation

Heterogeneous domain adaptation (HDA) addresses the chal-lenge of domain adaptation in scenarios where the source and target domains exhibit distinct types of features. While nu-merous studies have considered the HDA problem and demon-strated efficacy in transferring knowledge across diverse feature types, existing methods are predominantly tailored for I-D feature vectors, leaving the handling of high-order features largely under-explored. In this paper, we address a new problem, namely, heterogeneous tensor domain adaptation (HTDA), where features from either the source or target do-mains are tensors with different sizes or orders. Our approach involves a multilinear principal component analysis (MPCA) based tensor projection method which projects heterogeneous features onto a shared latent tensor space. Additionally, we leverage a probabilistic class-wise distribution alignment to align feature distributions in the tensor space, followed by label propagation from both the source data and labeled target data to unlabeled target data through graph regularization. Our experimental results on multiple domain adaptation datasets demonstrate the superior performance of our proposed tensor-based method in comparison to existing vector-based HDA methods.