Guided Manifold Alignment With Geometry-Regularized Twin Autoencoders

Jake S. Rhodes, Adam G. Rustad, Marshall S. Nielsen, Morgan Chase McClellan, Dallan Gardner, D. W. Hedges

TL;DR

This paper introduces a novel "guided manifold alignment" framework using geometry-regularized twin autoencoders (AEs) to overcome the out-of-sample extension limitation of traditional manifold alignment methods. By leveraging pre-aligned embeddings and a multitask learning approach, their method enhances cross-domain generalization and enables accurate prediction on unseen data, as demonstrated in Alzheimer's disease diagnosis.

What Problem Does This Paper Address?

Manifold alignment (MA) aims to find a shared, low-dimensional representation for data from different sources (domains or modalities). This is incredibly useful for tasks like data visualization, dimensionality reduction, and integrating diverse datasets. However, a major hurdle for many existing MA techniques is their inability to handle "out-of-sample" data – meaning they can't easily predict the representation for new data points that weren't part of the original training set. This severely limits their real-world applicability, as retraining the entire model for every new data point is often impractical or impossible. This paper tackles this limitation by developing a method that not only aligns manifolds but also generalizes well to unseen data.

Methods

The core of the proposed solution is a "guided representation learning framework" built upon geometry-regularized twin autoencoders (AEs). Here's a breakdown:

  • Twin Autoencoders: Two AEs are used, one for each domain (let's call them Domain X and Domain Y). Each AE consists of an encoder (mapping data to a shared latent space) and a decoder (reconstructing data from the latent space).
  • Geometry Regularization: The AEs are trained not just to reconstruct the input data but also to ensure that the encoded representations in the shared latent space preserve the geometric structure of the original data. This is achieved through a combined loss function that includes reconstruction loss and an alignment loss.
  • Guided Alignment: Crucially, the encoders are "guided" by a pre-trained alignment model. This means the AEs are encouraged to map data points to a latent space that is already aligned according to a traditional MA method. This provides a strong prior for the alignment.
  • Multitask Learning: The framework uses a multitask learning formulation. The AEs simultaneously learn to reconstruct data and to produce aligned embeddings. This joint training helps improve robustness and generalization.
  • Out-of-Sample Extension: Once trained, the twin AEs can take new data points from either domain, encode them into the shared latent space, and then use the decoders to reconstruct them or map them to the other domain. The decoder-swapping capability allows for cross-domain translation.

Datasets

The paper evaluates its approach on several benchmark datasets from the UCI Machine Learning Repository. These datasets are used for both classification and continuous response tasks. Specific examples include:

  • Classification: Artificial Tree, Balance Scale, Breast Cancer, Car, Chess, CRX, Diabetes, Glass, Heart Disease, Hepatitis, Ionosphere, Iris, Optical Digits, Parkinson's, Seeds, Tic-Tac-Toe, Wine.
  • Continuous Response: Airfoil Self-Noise, Air Quality, Automobile, Auto MPG, Beijing PM25, Community Crime, Concrete Compressive Strength, Concrete Slump Test, Cycle Power Plant, Energy Efficiency, Facebook Metrics, Five Cities PM25, Hydrodynamics, Istanbul Stock, Naval Propulsion Plants, Optical Network, Parkinsons, Protein, SML 2010.

Additionally, the paper applies its framework to a real-world problem using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), specifically combining ADAS-Cog 13 and FAQ scores.

Key Insights and Ideas

  • Pre-aligned Embeddings as a "Guide": Using an existing, albeit potentially limited, alignment as a "ground truth" for the AE's encoder is a powerful way to steer the representation learning process. This is more effective than just relying on reconstruction alone.
  • Twin AEs for Cross-Modal Fidelity: The twin AE architecture, with its shared latent space and cross-modal mapping capabilities (via decoder swapping), is a natural fit for learning representations that are consistent across different data modalities.
  • Multitask Learning for Robustness: Combining reconstruction and alignment objectives within the AE framework significantly improves the model's ability to generalize and handle noisy or incomplete data.
  • Practical Out-of-Sample Extension: The proposed method provides a principled way to extend MA to new data, which is a critical step for deploying these techniques in real-world scenarios where data is continuously generated.
  • Alzheimer's Diagnosis Application: The successful application to ADNI data highlights the potential of this method to integrate disparate clinical and cognitive measures, potentially leading to earlier and more accurate diagnoses, especially when one type of data is scarce.

Key Results

  • Improved Embedding Consistency: The twin AE framework, when regularized by various MA methods, shows significantly higher Mantel test correlations (a measure of embedding similarity) compared to using the original MA embeddings alone, especially when the alignment loss is active (λ > 0). For example, with JLMA, correlations reach 0.80 ± 0.32.
  • Enhanced Predictive Performance: In a k-NN based evaluation, the AE-aligned representations consistently outperformed baseline models trained on individual domains for most datasets, demonstrating that the shared representation captures valuable information. Methods like MASH, SPUD, and MAGAN as regularizers showed particularly strong improvements.
  • Superior Cross-Domain Mapping: The twin AE's decoder-swapping mechanism achieved lower Mean Squared Error (MSE) for cross-domain mapping compared to competing methods like MAGAN, DTA, and MASH, indicating more accurate translation of data points between domains.
  • Alzheimer's Disease Application: The framework successfully translated between ADAS-Cog 13 and FAQ scores, demonstrating its utility in integrating multi-modal patient data. The RMSE values for translation were generally low, suggesting good predictive accuracy.

Limitations and Future Work

  • Dependence on Initial Alignment: The performance of the guided MA framework is influenced by the quality of the initial alignment model used for guidance. While the AE improves upon it, a poor initial alignment might still limit the final outcome.
  • Scalability for N > 2 Domains: The current framework is described and tested for two domains. While the authors state it's generalizable to N domains, experimentation on multiple domains is deferred to future work.
  • Anchor Point Sensitivity: The paper notes that while anchor points are crucial for alignment, the specific influence of the anchor loss coefficient (λ) was found to be negligible after initial alignment, suggesting that the underlying MA model's ability to establish anchors is more critical.
  • Future Work: The authors suggest extending the framework to handle more than two domains and exploring its application in more complex, real-world scenarios where data imbalance and noise are significant challenges.

Summary updated

Abstract

Manifold alignment (MA) involves a set of techniques for learning shared representations across domains, yet many traditional MA methods are incapable of performing out-of-sample extension, limiting their real-world applicability. We propose a guided representation learning framework leveraging a geometry-regularized twin autoencoder (AE) architecture to enhance MA while enabling generalization to unseen data. Our method enforces structured cross-modal mappings to maintain geometric fidelity in learned embeddings. By incorporating a pretrained alignment model and a multitask learning formulation, we improve cross-domain generalization and representation robustness while maintaining alignment fidelity. We evaluate our approach using several MA methods, showing improvements in embedding consistency, information preservation, and crossdomain transfer. Additionally, we apply our framework to Alzheimer's disease diagnosis, demonstrating its ability to integrate multi-modal patient data and enhance predictive accuracy in cases limited to a single domain by leveraging insights from the multi-modal problem.
BibTeX
@misc{rhodes2025guided,
  author = {Jake S. Rhodes and Adam G. Rustad and Marshall S. Nielsen and Morgan Chase McClellan and Dallan Gardner and D. W. Hedges},
  title = {{Guided Manifold Alignment With Geometry-Regularized Twin Autoencoders}},
  year = {2025},
  eprint = {2509.22913},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.22913},
}
Figure