Generalized Robust Adaptive-Bandwidth Multi-View Manifold Learning in High Dimensions with Noise

Multiview datasets are common in scientific and engineering applications, yet existing fusion methods offer limited theoretical guarantees, particularly in the presence of heterogeneous and high-dimensional noise. We propose Generalized Robust Adaptive-Bandwidth Multiview Diffusion Maps (GRAB-MDM), a new kernel-based diffusion geometry framework for integrating multiple noisy data sources. The key innovation of GRAB-MDM is a {view}-dependent bandwidth selection strategy that adapts to the geometry and noise level of each view, enabling a stable and principled construction of multiview diffusion operators. Under a common-manifold model, we establish asymptotic convergence results and show that the adaptive bandwidths lead to provably robust recovery of the shared intrinsic structure, even when noise levels and sensor dimensions differ across views. Numerical experiments demonstrate that GRAB-MDM significantly improves robustness and embedding quality compared with fixed-bandwidth and equal-bandwidth baselines, and usually outperform existing algorithms. The proposed framework offers a practical and theoretically grounded solution for multiview sensor fusion in high-dimensional noisy environments.

💡 Research Summary

The paper addresses the problem of fusing multiple high‑dimensional data sources (views) that are observed through heterogeneous sensors and corrupted by different levels and structures of noise. Existing multiview fusion methods, such as multiview diffusion maps (MDM), canonical correlation analysis (CCA), kernel CCA, and NCCA, either assume a common bandwidth for all views or rely on heuristic parameter choices, which leads to poor performance when the views have disparate dimensions or noise characteristics.

To overcome these limitations, the authors propose Generalized Robust Adaptive‑Bandwidth Multiview Diffusion Maps (GRAB‑MDM). The core idea is a view‑dependent bandwidth selection procedure that automatically adapts to the geometry and noise level of each view. The procedure consists of two stages: (i) a view‑specific scale is estimated from local distances (e.g., median or mean of pairwise distances) to capture the intrinsic signal magnitude and noise variance of that sensor; (ii) a global scaling factor is obtained by averaging (or weighted averaging) the view‑specific scales across all views. The final bandwidth for view ℓ is εℓ = γ·sℓ, where γ is the global factor and sℓ is the view‑specific scale. This adaptive choice ensures that noisy, high‑dimensional views receive a larger bandwidth (more smoothing), while cleaner, lower‑dimensional views use a smaller bandwidth (preserving fine structure).

Using the adaptive bandwidths, the algorithm constructs a kernel matrix Kℓ for each view ℓ with entries Kℓ(i,j)=K(‖yℓi−yℓj‖²/εℓ), where K is a standard positive‑definite kernel such as the Gaussian. Cross‑view affinity matrices are then formed by matrix multiplication Kℓ1,ℓ2 = Kℓ1 Kℓ2, which effectively propagates shared manifold information across views. All cross‑view blocks are assembled into a block‑affinity matrix K of size (K·n)×(K·n). After degree normalization (D = diag(K·1) ), a row‑stochastic transition matrix A = D⁻¹K is obtained. Because A is similar to a symmetric matrix, its eigen‑decomposition yields eigenvalues η₁≥η₂≥… and right eigenvectors u₁,u₂,… . The embedding is built from the first m nontrivial eigenpairs, with diffusion time t typically set to 1. The authors also propose an elbow‑based automatic selection of m, incorporating eigenvalue ratios to make the method fully data‑driven.

The theoretical contribution rests on a common‑manifold model: each clean signal xℓi lies on a smooth, compact d‑dimensional manifold M, and each sensor observes a diffeomorphic copy of M embedded in its own ambient space ℝ^{pℓ}. Noisy observations are yℓi = xℓi + ξℓi, where ξℓi are independent sub‑Gaussian noise vectors with possibly different variances across views. Under this model, the authors prove that the normalized graph Laplacian built from the adaptive kernels converges, as n → ∞ and εℓ → 0 at appropriate rates, to a mixture of Laplace‑Beltrami operators on the view‑specific embedded manifolds plus lower‑order bias terms (Definition 3.3). The bias is shown to be O(ε) and the variance O((n ε^{d/2})⁻¹) (Theorems 4.2 and 4.4), matching classical diffusion‑map rates but now holding uniformly across heterogeneous views. In Section 5 they further demonstrate that when the signal‑to‑noise ratio (SNR) is sufficiently high, the leading noise term cancels out, yielding a limiting operator that is robust to the heterogeneous noise. This provides a rigorous justification for the empirical robustness observed in experiments.

Computationally, the block structure of K enables an efficient implementation. Each block Kℓ can be processed independently, and the full eigen‑decomposition of A can be reduced to singular‑value decompositions of the normalized blocks, dramatically lowering memory and runtime costs for large K and n. Moreover, the transition matrix is designed to allow only cross‑view moves, suppressing “lazy walks” that stay within a single view; this design, inspired by prior work on diffusion maps, further improves noise robustness (Remark 2.2).

Extensive experiments are conducted on synthetic data (where each view is assigned different dimensions and noise levels) and on real multimodal datasets such as multispectral imagery and motion‑capture recordings. Baselines include fixed‑bandwidth MDM, equal‑bandwidth MDM, Alternating Diffusion Maps (ADM), NCCA, and deep‑learning fusion models. Evaluation metrics comprise clustering accuracy, normalized mutual information, and visual inspection of low‑dimensional embeddings. GRAB‑MDM consistently outperforms baselines, especially when noise heterogeneity is pronounced. The adaptive bandwidths automatically balance smoothing across views, leading to clearer cluster separation and more faithful preservation of the underlying manifold geometry.

In summary, the paper makes three major contributions: (1) a principled, data‑driven view‑dependent bandwidth selection that adapts to geometry and noise; (2) a rigorous convergence and robustness analysis under a common‑manifold assumption with heterogeneous high‑dimensional noise; (3) an efficient block‑SVD implementation and extensive empirical validation. The work provides a solid theoretical foundation for multiview sensor fusion in challenging noisy environments and opens avenues for future research on non‑Euclidean views, limited‑sample regimes, and scalable approximations.