Spatiotemporal Augmentation via Graph Wavelets

• The paper introduces a robust framework using graph wavelet transforms to perturb, enrich, and regularize spatiotemporal data on dynamic graphs.
• It employs fast Chebyshev polynomial approximations and dynamic kernel designs to efficiently compute multi-scale wavelet transforms without costly eigendecomposition.
• Applications demonstrate significant improvements in dynamic GNN regularization, event localization, and discovery of latent multiscale patterns across diverse domains.

Spatiotemporal Augmentation via Graph Wavelets refers to a family of techniques rooted in spectral graph theory that leverage graph wavelet transforms to perturb, enrich, or regularize data defined on graphs that evolve over space and time. These methods have emerged as essential tools for machine learning on dynamic graphs, providing principled ways to manipulate signals and adjacency structures while respecting both the underlying spatial topology and temporal dynamics. Applications include improving the robustness and generalization of dynamic graph neural networks (GNNs), denoising network signals, localization of spatiotemporal events, and unveiling multiscale latent patterns.

1. Foundations: Spectral Graph Wavelets and Spatiotemporal Graphs

Let $G=(V,E,w)$ be a finite weighted graph, with adjacency $A\in\mathbb{R}^{N\times N}$ and (combinatorial) Laplacian $L=D-A$. The spectral decomposition $L=U\Lambda U^\top$ provides the basis for defining graph Fourier and wavelet transforms. For a wavelet-generating kernel $g:\mathbb{R}^+\to\mathbb{R}$, the wavelet operator at scale $t>0$ is $T^t_g = g(tL) = U g(t\Lambda) U^\top$, enabling construction of wavelets via $\psi_{t,n}=T^t_g\delta_n$ (0912.3848).

To extend this to the spatiotemporal regime, the product-graph $\widetilde{G}$ is formed, whose vertices are $(v,t)$ with $v\in V$ and $t\in\{1,\dots,T\}$, and whose edges include both spatial and temporal couplings (e.g., spatial adjacency replicated across time-layers, and temporal links connecting $(v,t)$ to $(v,t+1)$) (Mendes et al., 2014, 0912.3848, Grassi et al., 2016). The Laplacian $\widetilde{L}$ encodes this joint structure and is used to define spatiotemporal wavelets.

2. Spatiotemporal Graph Wavelet Construction and Localization

The core construction involves defining wavelets adapted to the graph’s (possibly time-varying) spectrum and localizing them both in the vertex and time domains. For static graphs, the spectral graph wavelet at node $n$ and scale $t$ is

$$\psi_{t,n}(m) = \sum_{\ell=0}^{N-1} g(t\lambda_\ell) u_\ell(n) u_\ell(m).$$

For a spatiotemporal graph with $S$ spatial nodes and $T$ time steps, the Laplacian $\widetilde{L}$ acts on the $ST$-dimensional vector $f$, and the same formalism applies: $\Psi_{t,n} = g(t\widetilde{L})\delta_n$ (0912.3848).

Dynamic graph wavelet (DGW) frames can be built for evolving graphs using kernels in the joint space–time spectral domain. The discrete wave equation on a joint graph–time Laplacian $\mathcal{L}$ leads to dynamic kernels with the form

$$\widehat{W}_s(\lambda,\omega) = \frac{1-e^{-(\beta+j\omega)}}{1-2e^{-(\beta+j\omega)}\cos\theta+e^{-2(\beta+j\omega)}},$$

where $\theta=\arccos(1-\frac{s\lambda}{2})$, $s$ is a scale parameter, and $\beta>0$ controls damping (Grassi et al., 2016). These kernels allow explicit modeling of propagation dynamics across the network.

Localization properties are governed by the choice of $g$: e.g., for $g(\lambda) \sim \lambda^K$ near 0, wavelets exhibit rapid spatial decay away from the center as $t\to 0$ (0912.3848). For spatiotemporal graphs, sufficiently regular $g$ yields joint localization in both domains.

3. Fast Algorithms and Computational Considerations

A major computational bottleneck in spectral methods is the eigendecomposition of large Laplacians. Fast Chebyshev polynomial approximations can bypass this, expanding $g(tL)$ or $g(t\widetilde{L})$ as low-degree polynomials of $L$, enabling efficient computation via recurrence with cost linear in the number of edges per scale. For product graphs, this allows scalable computation of multi-scale, spatiotemporal wavelet transforms on large graphs (0912.3848).

For dynamic graphs composed of a sequence of temporal snapshots, such as $G_1, G_2, \dotsc, G_T$, techniques may involve incremental eigensolvers or leveraging block structure to accelerate computation. Random walk–based augmentations, such as the “SpatioTemporal Activity-Aware Random Walk Diffusion” (STAA), combine localized spectral analysis with random-walk diffusion to generate augmented adjacency matrices for each snapshot (Chu et al., 17 Jan 2025).

4. Augmentation Strategies Utilizing Graph Wavelets

Spatiotemporal augmentation with graph wavelets proceeds by injecting stochastically structured perturbations into the data or the adjacency structure, specifically leveraging the multi-scale, localized, and temporally aware nature of graph wavelets:

• Spectral perturbation: Additive or multiplicative modification of original signals $f$ with filtered $$\Delta f = \sum_{j\in\mathcal{S}}\alpha_j T_g^{t_j} f$$, where the subset $\mathcal{S}$ and coefficients $\alpha_j$ are randomly selected (0912.3848).
• Localized “bumps”: Injection of sparse, spatially and temporally localized wavelet atoms at selected nodes and time-points (e.g., adding $\beta_r \psi_{t_j,n_r}$ for randomly chosen $(t_j,n_r)$) (0912.3848).
• Adjacency augmentation: Construction of diffusion-based augmented adjacency matrices by guiding random walks with spatiotemporal metrics derived from wavelet coefficients, as in STAA, where activity-aware coefficients $\beta_{t,j}$ modulate the transition probabilities of random walks, emphasizing nodes with high spatiotemporal dynamics and suppressing noisy edges (Chu et al., 17 Jan 2025).

5. Practical Applications and Empirical Results

Graph wavelet–based spatiotemporal augmentation has enabled advances in several domains:

• Dynamic GNN regularization: By replacing raw adjacency matrices $A_t$ with augmented $X_t$ derived via wavelet-informed random walks, downstream GNNs exhibit improved robustness and generalization. STAA achieved state-of-the-art AUC and Macro F1 on dynamic link-prediction and node-classification benchmarks, outperforming DropEdge, GDC, and other augmentation baselines; for instance, AUC on BitcoinAlpha increased from $57.0\pm1.2$ (no augmentation, GCN) to $79.6\pm0.8$ with STAA (Chu et al., 17 Jan 2025).
• Source/epicenter localization: DGW frames enable sparse recovery and localization of spatiotemporal events—e.g., robustly localizing earthquake epicenters on seismic networks, with mean error $<40$ km at $0$ dB SNR (Grassi et al., 2016).
• Multiscale feature extraction: Time-augmented graph wavelets and tomograms enable discovery of latent multiscale patterns in financial, ecological, and connectomics networks (Mendes et al., 2014).

6. Generalizations and Further Extensions

Graph wavelet methodology generalizes to a broad class of dynamics by replacing the kernel $g$ or $\widehat{W}_s$ to model different processes:

• Diffusion dynamics: Use $g(\lambda)=e^{-s\lambda}$ for heat-type propagation.
• Advection-reaction and non-symmetric graphs: Modify the Laplacian or kernel appropriately.
• Time-varying and multilayer graphs: The Laplacian framework accommodates directed edges, time-dependent weights, and hypergraph or multi-relational extensions via Kronecker sums, allowing joint wavelet analysis in richer spatiotemporal regimes (Grassi et al., 2016).
• Learned spectral kernels: Parameterize $g$ or $\widehat{W}_s$ with learnable parameters, fitting dynamics from data via dictionary learning or within end-to-end neural frameworks (Grassi et al., 2016); in the context of dynamic GNNs, variants now allow gating parameters (e.g., $\gamma$, $\delta$) to be optimized with task losses (Chu et al., 17 Jan 2025).

7. Stability, Limitations, and Outlook

Frame stability is ensured if the lower frame bound $A>0$ for the collection $\{\psi_{t_j,n}\}$; choice of scales and kernel shapes must balance localization with coverage of the relevant frequency bands (0912.3848, Grassi et al., 2016). Potential limitations include the eigen-computation cost for large graphs (mitigated by polynomial acceleration), choice of graph signals for wavelet analysis, and hyperparameter selection. Ongoing extensions address continuous-time models, automatic parameter learning, and application to graphs with richer node-attribute and edge-feature contexts.

A plausible implication is that spatiotemporal augmentation via graph wavelets integrates mathematically principled multiscale, topology- and time-aware priors into graph learning pipelines, expanding both the theoretical foundations and practical toolsets for dynamic graph analysis (0912.3848, Grassi et al., 2016, Mendes et al., 2014, Chu et al., 17 Jan 2025).