Supervised Manifold Alignment

• Supervised Manifold Alignment is a technique that constructs unified low-dimensional embeddings from multi-domain data by integrating class labels and anchor correspondences.
• It employs methods like random forest–based graph construction, optimal transport, and twin autoencoders to preserve intrinsic geometry while aligning cross-domain features.
• Empirical results show that SMA improves cross-domain classification accuracy and adapts efficiently to multi-modal datasets under various computational constraints.

Supervised Manifold Alignment (SMA) refers to a class of algorithms designed to construct a shared, low-dimensional embedding from data collected across multiple domains or modalities, leveraging both geometric structure and explicit supervision, such as class labels or known cross-domain correspondences. The objective is to find an embedding in which related samples from each domain are brought into alignment, while preserving intrinsic manifold geometry within domains. SMA enables cross-domain learning, domain adaptation, and improved downstream inference in multi-view and multi-modal scenarios.

1. Problem Formulation and Theoretical Foundations

Let $\mathcal{X}=\{x_i\}_{i=1}^{n_s}$, $x_i \in \mathbb{R}^{d_s}$, and $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$, $y_j \in \mathbb{R}^{d_t}$, denote two domains. Each point may bear a class label ($y_i^s$, $y_j^t$) or be associated through anchor correspondences. The challenge is to find a joint embedding $Z = [Z^s; Z^t] \in \mathbb{R}^{(n_s+n_t)\times m}$ such that:

• Local geometry is preserved within each domain.
• Cross-domain correspondences and/or class structure are respected.

A canonical optimization problem is

$$\min_Z \operatorname{Tr}(Z^\top L Z), \quad \text{subject to } Z^\top D Z = I,$$

where $W$ encodes within- and cross-domain affinities, $D = \text{diag}(W\mathbf{1})$, and $x_i \in \mathbb{R}^{d_s}$0. The embedding $x_i \in \mathbb{R}^{d_s}$1 consists of the $x_i \in \mathbb{R}^{d_s}$2 eigenvectors with the smallest nonzero generalized eigenvalues of $x_i \in \mathbb{R}^{d_s}$3 (Rhodes et al., 2024).

2. Algorithmic Approaches and Methodological Variants

Random Forest–Supervised Graph Construction

Recent SMA methods leverage class supervision by constructing geometry-preserving affinities using random forests (RFs). Given class-labeled data in each domain, an RF trained on $x_i \in \mathbb{R}^{d_s}$4 or $x_i \in \mathbb{R}^{d_s}$5 induces a stochastic proximity matrix, where the proximity between samples $x_i \in \mathbb{R}^{d_s}$6 and $x_i \in \mathbb{R}^{d_s}$7 is

$x_i \in \mathbb{R}^{d_s}$8

with $x_i \in \mathbb{R}^{d_s}$9 trees (Rhodes et al., 2024).

RF-GAP proximities further employ out-of-bag observations to ensure robustness:

$\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$0

resulting in a row-stochastic affinity.

For cross-domain alignment, known anchor pairs are assigned strong cross-domain affinities, either as maximal within-domain proximities or as fixed positive weights. The full block adjacency

$\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$1

supports joint spectral embedding. This methodology empirically improves cross-domain classification, especially when class label structure is leveraged in proximity construction (Rhodes et al., 2024).

Probabilistic, Optimal Transport, and Latent Variable Models

Alternative SMA frameworks include probabilistic latent variable approaches (e.g., Manifold Alignment Determination—MAD (Damianou et al., 2017)), which place a prior over a latent space, fit Gaussian process mappings from latent to observed data in each domain, and infer alignments via variational inference. Small sets of aligned examples serve as priors for the model.

Other approaches, such as Diffusion Transport Alignment (DTA) (Duque et al., 2022), construct diffusion operators within each domain, propagate probability mass through anchor correspondences, and solve a partial optimal transport program over a diffusion-based inter-domain cost. Label supervision can be incorporated by modifying the cost matrix to increase the distances between points with distinct labels.

Twin autoencoder architectures (Rhodes et al., 26 Sep 2025) regularize parametric mappings (encoders/decoders) using a precomputed aligned embedding and anchor constraints, enabling out-of-sample extension and robust multi-modal translation.

3. Supervision Mechanisms

Supervised manifold alignment capitalizes on available side information in several forms:

• Class label integration (“supervised initialization”): Random forests incorporate label structure into proximities by training on class labels, optionally further amplifying same-class affinities (Rhodes et al., 2024).
• Anchor-based cross-domain constraints: Known alignments (anchor pairs) induce blockwise affinities or anchor loss terms, enforcing proximity in the embedding across domains (Rhodes et al., 2024, Duque et al., 2022, Rhodes et al., 26 Sep 2025).
• Label-based cost modifications: In diffusion transport methods, label disagreements are penalized directly in inter-domain cost matrices, e.g., $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$2 (Duque et al., 2022).
• Guided embedding alignment: Precomputed manifold alignments (from unsupervised or semi-supervised methods) serve as soft targets for supervised autoencoder guides (Rhodes et al., 26 Sep 2025).

The degree and type of supervision (labels, anchors, pre-alignments) direct the optimization landscape, influencing robustness and downstream fidelity.

4. Computational Complexity and Scalability

Complexity in SMA typically arises from three stages:

• Affinity/proximity computation: For random forest–based methods, training costs $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$3 (for $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$4 samples, $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$5 features, $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$6 trees) and proximity extraction $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$7 per domain (Rhodes et al., 2024).
• Eigen-decomposition: Joint Laplacian embedding for $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$8 points requires $\mathcal{Y}=\{y_j\}_{j=1}^{n_t}$9 operations. For practical regimes ($y_j \in \mathbb{R}^{d_t}$0), direct diagonalization is feasible; for large-scale settings, Nyström or landmark-based approximations become necessary (Rhodes et al., 2024).
• Probabilistic/variational inference: Complexity depends on the number of anchors ($y_j \in \mathbb{R}^{d_t}$1 per iteration for $y_j \in \mathbb{R}^{d_t}$2 inducing points in MAD), with matching costs determined by the size of the unaligned set (Damianou et al., 2017).
• Optimal transport: Sinkhorn iterations provide efficient convex solvers with convergence guarantees for partial transport in DTA (Duque et al., 2022).

The suitability of a method depends on computational resources, problem scale, anchor density, and real-time constraints.

5. Empirical Performance and Benchmarks

SMA techniques demonstrate substantive gains over unsupervised and single-domain methods in diverse benchmarks:

• On 16 UCI datasets with various splits (random, importance-based, distortion, rotation), random forest–supervised graph alignment (RF-MASH, RF-SPUD) outperformed all competitors, exceeding both domain-specific baseline classifiers in 48% and 45% of settings, respectively. Other approaches such as KEMA, RF-MALI, and DTA performed at 42%, 39%, and 35% (Rhodes et al., 2024).
• In MNIST-Double and real single-cell datasets, DTA achieved the lowest mean squared error in inter-domain regression on 8/9 tasks and the highest domain adaptation accuracy in 80% of scenarios (Duque et al., 2022).
• MAD achieved uniform or superior alignment accuracy (Kendall-τ scores on toy and real data) compared to linear and spectral alternatives with far fewer anchor correspondences, reflecting robustness to noise and private variation (Damianou et al., 2017).
• Geometry-regularized twin autoencoders preserved the geometric fidelity of backbone alignments (Mantel $y_j \in \mathbb{R}^{d_t}$3–$y_j \in \mathbb{R}^{d_t}$4 for JLMA, SPUD, MASH) and achieved superior downstream transfer learning performance, including in clinical applications (Rhodes et al., 26 Sep 2025).

Method	Win Rate over Baselines
RF-MASH	0.48
RF-SPUD	0.45
KEMA	0.42
RF-MALI	0.39
DTA	0.35
others	<0.30

6. Limitations, Practical Guidance, and Research Directions

SMA methods that exploit class structure and anchors offer strong empirical performance but exhibit several limitations:

• Memory and time complexity: $y_j \in \mathbb{R}^{d_t}$5 storage for affinity matrices and $y_j \in \mathbb{R}^{d_t}$6 eigensolvers constrain direct methods to modest sample sizes. Approximations (e.g., Nyström) are suggested for scalability (Rhodes et al., 2024).
• Dependence on supervision density: Performance is sensitive to the fraction and quality of anchors; insufficient or noisy supervision may compromise alignment robustness (Duque et al., 2022, Rhodes et al., 2024).
• Model fit: Random forest proximities and guided autoencoders rely on effective training of foundational models and may struggle with class imbalance or label noise (Rhodes et al., 2024).
• Choice of hyperparameters: Embedding dimension $y_j \in \mathbb{R}^{d_t}$7, number of trees $y_j \in \mathbb{R}^{d_t}$8, anchor proportion, and regularization balance ($y_j \in \mathbb{R}^{d_t}$9) require tuning, often via downstream validation (Rhodes et al., 2024, Rhodes et al., 26 Sep 2025).

Open research problems include:

• Scalable, streaming, and online SMA.
• Alternative supervised proximity models (e.g., gradient-boosted trees, neural nets).
• Rigorous error bounds relating anchor density and manifold complexity.
• Theoretical guarantees and sharp generalization results for parametric extension architectures.

A plausible implication is that as complex, multi-modal datasets proliferate, SMA frameworks—especially those integrating class structure and robust cross-domain geometry—will play a central role in domain adaptation and multimodal inference due to their ability to exploit limited supervision efficiently (Rhodes et al., 2024, Duque et al., 2022, Rhodes et al., 26 Sep 2025).