Unfolded Laplacian Spectral Embedding (ULSE)

• The paper introduces ULSE, a framework that aggregates spectral information from multiple time snapshots of normalized Laplacians, ensuring both cross-sectional and longitudinal stability.
• ULSE computes a time-aware truncated SVD on the unfolded operator, linking latent community structures to global cut-based metrics and Cheeger-style bounds.
• Empirical results on synthetic and real-world networks show ULSE’s improved clustering accuracy, normalized mutual information, adjusted Rand index, and F1 scores compared to prior methods.

Unfolded Laplacian Spectral Embedding (ULSE) is a method for representing nodes in dynamic networks by leveraging the joint spectral properties of sequences of normalized graph Laplacians. ULSE generalizes earlier unfolded adjacency spectral embedding frameworks, ensuring stable and interpretable embeddings that satisfy rigorous cross-sectional and longitudinal stability guarantees in time-evolving graphs. By aggregating the spectral information across multiple time steps, ULSE provides a consistent representation that is robust to temporal fluctuations, directly linking latent community structure to global and temporal cut-based measures.

1. Mathematical and Algorithmic Foundations

ULSE operates on a dynamic network comprising T graph snapshots, each represented by an adjacency matrix $A^{(t)}$ and corresponding degree matrix $D^{(t)}$ for $t = 1,\ldots,T$. For each snapshot, ULSE constructs the normalized Laplacian: $$L^{(t)} = I - (D^{(t)})^{-1/2} A^{(t)} (D^{(t)})^{-1/2}.$$ The method then horizontally concatenates (or “unfolds”) the sequence $\{L^{(t)}\}_{t=1}^T$ to produce an aggregate operator, on which a truncated singular value decomposition (SVD) is performed. The main variants of ULSE differ in the construction of the unfolded operator:

• ULSE–n1: Concatenate $L^{(t)}$ for all $t$ and retain the $d=K-1$ lowest nontrivial singular values for embedding.
• ULSE–n2: Use a partially time-aggregated degree normalization by forming $D^{(1:T)}$ as the sum of all degree matrices and set

$$L^{(t)} = - (D^{(1:T)})^{-1/2} A^{(t)} (D^{(t)})^{-1/2},$$

with embedding dimension $d=K$ (the rank of the latent block structure).

After SVD, embeddings $Y^{(t)}$ for nodes at time $t$ are extracted and corrected (aligned) to ensure consistency, typically by centering using an anchor subspace $U$. This process stitches instantaneous and persistent latent structure into the node representations.

2. Stability Guarantees: Cross-Sectional and Longitudinal

A central contribution of ULSE is its provable satisfaction of two stability properties crucial for dynamic representation:

• Cross-sectional stability: For any $i,j$ and fixed $t$, if nodes $i$ and $j$ have identical rows in the probability matrix $P^{(t)}$ (i.e., identical latent behaviors at time $t$), then $Y_{i}^{(t)} = Y_{j}^{(t)}$.
• Longitudinal stability: For any node $i$ and times $t, u$, if $P_{i:}^{(t)} = P_{i:}^{(u)}$, then $Y_i^{(t)} = Y_i^{(u)}$.

These properties are established first in the population (noise-free) setting, and then extended to the observed data via spectral norm perturbation bounds: $$\max_{i,t}\, \| Y_{i}^{(t)} - \widetilde{Y}_{i}^{(t)} \| = O\left( \frac{1}{\rho^{1/2} n^{1/2}} \right) \quad \text{a.s.}$$ with $\rho$ a sparsity parameter and $n$ the number of nodes.

The proof uses deviation inequalities for the spectral norm between the sample and population unfolded Laplacians, along with Davis–Kahan-type subspace perturbation results for the left singular vector subspaces.

3. Cheeger-Style Inequality and Conductance Connection

ULSE introduces a Cheeger-style inequality linking the singular values of the unfolded normalized Laplacian to the $k$-way conductance of the dynamic graph: $\phi_k(G) = \max_{t \in [T]} \phi_k(G^{(t)}).$ The $k$-th singular value $\sigma_k$ of the unfolded operator satisfies

$$\frac{\sqrt{\max\{\sigma_k^2 - \min_t \|L^{-t}\|^2, 0\}}}{2} \leq \phi_k(G) \leq \mathrm{poly}(k) \sqrt{\sigma_k},$$

where $L^{-t}$ denotes the operator with the $t$-th snapshot removed. This result extends classical Cheeger and higher-order Cheeger inequalities to the unfolded setting, connecting the spectrum of the embedding operator to temporal cut and partition properties in the dynamic network.

4. Empirical Results and Benchmarks

Empirical evaluation on synthetic dynamic stochastic block models and real-world datasets such as brain connectivity graphs, school interaction networks, and stock market networks confirms the theoretical findings:

• Cross-sectional stability: Nodes assigned to the same block at a fixed time have near-identical embeddings.
• Longitudinal stability: Node embeddings for unchanged latent behavior over time remain stable.
• Performance metrics: ULSE achieves higher clustering accuracy, normalized mutual information, adjusted Rand index, and F1 scores compared to unfolded adjacency spectral embedding (UASE) and deep dynamic graph learning architectures (Node2Vec, JODIE, TGN), particularly in regimes with node sparsity or block merging.
• Visualization: t–SNE plots of the embeddings demonstrate preservation of both temporal persistence and dynamic community transitions.

5. Methodological Innovations and Variants

A distinctive feature of ULSE is the systematic replacement of the adjacency operator with the normalized Laplacian prior to unfolding. This modification improves interpretability and naturally leads to embeddings that reflect both degree normalization and cut-based properties. By controlling time-dependent normalization (in ULSE–n2) versus snapshot-level normalization (in ULSE–n1), the framework allows fine-tuning for different data characteristics (e.g., degree heterogeneity, rate of temporal graph change).

The result is a representation that robustly encodes long-term node behaviors, evolving community structure, and cut-based global properties, all while rigorously controlling perturbation-induced deviation from the population-level embeddings.

6. Theoretical Implications and Connections

Theoretical advances in ULSE extend the understanding of spectral graph embeddings by connecting the unfolding process (across time or relations) with normalized Laplacian operator theory. The stability conditions enforced by ULSE have direct analogues in spectral convergence results for dynamic random graph models; the Cheeger-style bounds extend the bridge between spectral gaps and combinatorial separation to the dynamic, multi-snapshot context. In this sense, ULSE situates itself at the intersection of temporal graph analysis, manifold learning, and operator-theoretic approaches to data representation.

7. Applications and Future Directions

ULSE is suited for scenarios in which dynamic network data require temporally consistent, interpretable node representations, including:

• Temporal community detection in social, biological, or financial networks;
• Anomaly detection and temporal forecasting, where stable normal behavior must be distinguished from transients;
• Knowledge graph and recommendation systems with time-dependent interactions.

The method’s theoretical grounding, explicit conductance connection, and robust empirical performance create a foundation for further extensions, such as incorporating side information, handling asynchronous or irregular time sampling, or combining with deep learning architectures for scalable inference in large-scale evolving graphs.

---

In summary, unfolded Laplacian spectral embedding establishes a principled, dynamically stable, and theoretically governed framework for time-evolving network representation, rigorously linking spectral structure to both local and global temporal graph properties and enabling consistent inference and analysis across heterogeneous dynamic network settings (Ezoe et al., 18 Aug 2025).