Tensor-Based Domain Alignment

• Tensor-Based Domain Alignment is a method that represents source and target data as higher-order tensors, enabling structured unsupervised adaptation.
• It employs mode-wise alignment, Tucker subspace projection, and Riemannian geometry to reduce domain discrepancies while preserving intrinsic features.
• Empirical results demonstrate improved accuracy and efficiency in applications such as visual recognition, hyperspectral imaging, and audio classification.

A tensor-based domain alignment (DA) algorithm is a methodology for unsupervised domain adaptation in which both source and target data are represented intrinsically as tensors—multi-way arrays or structured objects—rather than as flattened vectors. By leveraging higher-order structure, these algorithms perform mode-wise transformations, subspace projections, and alignment operations to minimize domain discrepancy while preserving essential features such as variance, local geometry, or Riemannian structure. Theoretical frameworks span Tucker decompositions, alignment matrices constrained on matrix manifolds (Stiefel or oblique), Riemannian geometry of the manifold of symmetric positive semi-definite (SPSD) tensors, and regularization schemes. This class includes a family of algorithms—TAISL, TA, TDA, and geometric approaches—which generalize and unify many linear subspace and covariance-based DA methods.

1. Tensor Representations and Problem Settings

In tensor-based DA, both source and target domains are comprised of $N_s$ source samples and $N_t$ target samples, each represented as $K$-th order tensors rather than vectors. For modes $n_1 \times \cdots \times n_K$, the source set is stacked as $$\mathcal{X}_s \in \mathbb{R}^{n_1 \times \cdots \times n_K \times N_s}$$ and the target as $$\mathcal{X}_t \in \mathbb{R}^{n_1 \times \cdots \times n_K \times N_t}$$. This abstraction naturally models structured data in computer vision and signal processing—e.g., CNN activation maps, hyperspectral cubes, audio time-frequency representations, or low-rank covariance matrices.

The fundamental problem is: given feature tensors from source and target domains with different marginal distributions, learn transformations (alignment matrices, projection subspaces) and ideally shared low-dimensional representations that (1) reduce cross-domain discrepancy and (2) retain the intrinsic, class-discriminative structure conserved across domains (Lu et al., 2017, Qin et al., 2018, Lee et al., 26 Jan 2026, Yair et al., 2020).

2. Alignment and Projection Frameworks

Core mechanisms underlying tensor-based DA include the following:

• Mode-wise Alignment Matrices: For each mode $k=1,\ldots,K$, introduce an alignment matrix $M^{(k)} \in \mathbb{R}^{n_k \times n_k}$, often constrained such that $M^{(k)}(M^{(k)})^T = I$ (Stiefel manifold) or $\operatorname{diag}(M^{(k)}(M^{(k)})^T)=I$ (oblique manifold) (Lu et al., 2017, Lee et al., 26 Jan 2026).
• Multilinear Tucker Subspace: A shared tensor subspace is sought in the form of factor matrices $U^{(k)} \in \mathbb{R}^{n_k \times d_k}$ satisfying $(U^{(k)})^TU^{(k)}=I$ for all $k$, along with domain-specific or shared core tensors (Lu et al., 2017, Qin et al., 2018, Lee et al., 26 Jan 2026).
• Tensor Alignment Operators: The domain-aligned source tensor is computed as $$[\mathcal{X}_s; \mathcal{M}] = \mathcal{X}_s \times_1 M^{(1)} \cdots \times_K M^{(K)}$$ and likewise for the target. Projection into the invariant subspace is then $$[\mathcal{G}_s; \mathcal{U}] = \mathcal{G}_s \times_1 U^{(1)} \cdots \times_K U^{(K)}$$.
• Riemannian Geometry for Covariance Tensors: For algorithms operating on sets of covariance descriptors, each data point is represented as a $d\times d$ SPSD matrix $C$, decomposed $C = GPG^T$ with $G$ on the Stiefel manifold and $P$ symmetric positive definite (SPD), facilitating alignment in both the subspace (Grassmann) and covariance (SPD) components (Yair et al., 2020).

This factorized approach enables flexible adaptation to diverse data types, including images, hyperspectral cubes, and time-series representations.

3. Optimization Objectives and Constraints

The learning objective typically unifies three components:

1. Alignment Loss / Source-Target Reconstruction: Penalizes the misfit between the domain-aligned, projected source (and optionally target) tensors and their lower-dimensional core representations:

$$\|\,[\mathcal{X}_s; \mathcal{M}] - [\mathcal{G}_s; \mathcal{U}]\,\|_F^2 + \|\mathcal{X}_t - [\mathcal{G}_t; \mathcal{U}]\|_F^2$$

or, in the most general form, sums over both domains after alignment and projection (Lu et al., 2017, Qin et al., 2018, Lee et al., 26 Jan 2026).

2. Variance-Preserving Regularizer: Enforces that the double-projected or reconstructed source (and possibly target) tensors still approximate the original, promoting preservation of intrinsic variance and preventing trivial solutions:

$$\lambda\,\|\,[\,[\mathcal{X}_s;\mathcal{M}];\mathcal{M}^T] - \mathcal{X}_s\,\|_F^2$$

or equivalently via trace expressions in the oblique TDA formalism (Lu et al., 2017, Lee et al., 26 Jan 2026).

1. Manifold Regularization/Graph Laplacians: Especially for spatially-structured data, additional Laplacian regularizers can enforce that local neighborhood structure is retained among projected core tensors in both domains (Qin et al., 2018).
2. Manifold Constraints: The alignment and projection matrices are typically constrained to be orthonormal (Stiefel) or unit-norm per row (oblique), encoding invariances and enabling efficient optimization on matrix manifolds (Lu et al., 2017, Lee et al., 26 Jan 2026).

4. Iterative Algorithmic Schemes

Tensor-based DA algorithms employ alternating or block-coordinate optimization:

• Subspace Update (U-step): With alignment matrices fixed, solve for the shared Tucker subspace and core tensors via higher-order SVD, alternating least squares, or convex decomposition of mode-wise unfoldings.
• Alignment Update (M-step): With subspace bases fixed, solve for each mode’s alignment matrix. This often reduces, per mode, to small-scale orthogonal Procrustes or Riemannian gradient steps with manifold retraction.
• Regularizer/Grap Laplacian Update: When Laplacian terms are included, compute spectral or nearest-neighbor graphs and close-form solutions for manifold-regularized core minimizations (particularly in hyperspectral and spatial settings).

For example, the TDA algorithm (Lee et al., 26 Jan 2026) utilizes a block-coordinate descent over alignment matrices ($M$) and subspace bases ($U$), with $M$ updated by Riemannian gradients on the oblique manifold and $U$ by Tucker ALS over SVDs on mode-wise unfoldings. Stopping criteria are based on relative decrease in global objective or iterate distance. Convergence is empirically fast—typically 20–30 iterations for modern oblique variants, versus 80+ for original TAISL (Lee et al., 26 Jan 2026, Lu et al., 2017). Manifold constraints guarantee feasible iterates throughout.

The SPSD-geometric DA (Yair et al., 2020) computes means and canonical representations on the Grassmann manifold and SPD cone, then uses parallel transport and exponential/logarithmic maps to re-align source covariance descriptors to the target mean, ensuring mean-matching but not necessarily isometry.

5. Special Cases, Generalizations, and Theoretical Guarantees

The TDA framework (Lee et al., 26 Jan 2026) explicitly unifies prior tensor-based methods via settings of its manifold constraint and regularization:

Alignment Constraint	Variance-Regularization	Method Recovered
Stiefel	Absent	TAISL (Lu et al., 2017)
Oblique	Absent	E-TAISL(O)
Stiefel/Oblique	Present (λ>0)	TDA(S)/TDA(O)

The canonical SPSD DA (Yair et al., 2020) employs exact mean matching in Grassmann/SPD factors, with convergence of mean calculations (Moakher, Absil) and cost dominated by SVD-based manifold computations. The TA algorithm for hyperspectral DA (Qin et al., 2018) is brought under the Tucker framework with Laplacian-regularized alternations and closed-form updates, provably converging to a stationary point in $\leq 15$ iterations.

All frameworks feature unsupervised learning—no target labels are used. For vectorized methods, tensor approaches avoid curse-of-dimensionality scaling by mode-wise adaptation.

6. Empirical Performance and Application Domains

Tensor-based DA algorithms have demonstrated superior accuracy and robustness across diverse data types:

• Visual Recognition: On Office–Caltech10 (conv feature stacks), TAISL achieves $86.9\%$ mean accuracy (vs. SA $74.3\%$, CORAL $72.0\%$, NTSL $85.3\%$) (Lu et al., 2017).
• Hyperspectral Image Classification: The TA and TA_P algorithms yield $83$–$86\%$ OA on Pavia U→C (vs. $67$–$75\%$ for non-tensor DA), with major gains in low-label regimes (Qin et al., 2018).
• Audio Scene Classification: TDA(O) achieves $85.7\%$ accuracy on TUT Urban Acoustic Scenes, outperforming all tested subspace and DA baselines (Lee et al., 26 Jan 2026).
• Cross-Domain Robustness: Tensor DA methods preserve class separability, decrease intra-class distances, and maintain high accuracy even with small or imbalanced target sets (Lu et al., 2017, Lee et al., 26 Jan 2026).

Convergence speed and runtime efficiency are enhanced by oblique manifold constraints, reducing iterations and compute cost compared to classical orthogonal approaches (Lee et al., 26 Jan 2026).

7. Role of Geometric Operators and Manifold Structures

Certain variants operate on SPSD matrix representations, introducing fine-grained geometric manipulations:

• SPSD Factorization: Each data point is mapped to $C = G P G^T$, separating subspace (Grassmann) and covariance (SPD) structure (Yair et al., 2020).
• Geodesic, Logarithmic, and Exponential Maps: Approximate geodesics, log, and exp maps are used for updating, transporting, and projecting SPSD factors on their respective manifolds.
• Parallel Transport: Explicit closed-form PT aligns tangent vectors (differences) in both geometric factors, enforcing mean alignment from source to target (Yair et al., 2020).
• Mean and Canonical Representations: Shared centers are computed in the Grassmann manifold (for orientation) and SPD cone (for scale/shape), aligning all data prior to transport.

This geometric design underpins the preservation of higher-order feature relationships implicit in source and target distributions.

---

References:

(Lu et al., 2017) "When Unsupervised Domain Adaptation Meets Tensor Representations" (Qin et al., 2018) "Tensor Alignment Based Domain Adaptation for Hyperspectral Image Classification" (Yair et al., 2020) "Symmetric Positive Semi-definite Riemannian Geometry with Application to Domain Adaptation" (Lee et al., 26 Jan 2026) "An Unsupervised Tensor-Based Domain Alignment"