Temporal Graph Clustering Methods

• Temporal Graph Clustering is a framework that partitions evolving graphs into groups with cohesive temporal and structural properties.
• Methodological approaches include spectral methods, probabilistic models, matrix factorization, and deep learning techniques to capture dynamic interactions.
• Key challenges involve enforcing causal constraints, managing batch processing tradeoffs, and adaptively determining cluster numbers amid sparse data.

Temporal Graph Clustering (TGC) encompasses the algorithmic and statistical frameworks for partitioning the nodes, edges, or snapshots of time-evolving or temporally annotated graphs into meaningful, temporally coherent clusters. TGC generalizes classical static graph clustering by explicitly incorporating the temporal dynamics of interactions, edge weights, or node features. This includes, but is not limited to, community detection in temporal networks, discovery of persistent interaction motifs, and joint segmentation of multivariate spatiotemporal data streams. Recent methodological innovations span spectral, probabilistic, deep learning, and matrix factorization paradigms, with applications ranging from social networks and climate science to neuroscience and record linkage.

1. Formal Definitions and Core Problem Settings

The canonical temporal graph is modeled as $G = (V, E, T)$, with $V$ the (possibly time-varying) node set, $E \subseteq V \times V \times T$ a set of time-stamped interactions, and $T$ the set or range of time indices or stamps. The temporal graph clustering problem is to partition $V$ (or $E$, or the snapshots themselves) into $K$ clusters $\{C_1, ..., C_K\}$, such that members of the same cluster exhibit similar temporal and structural properties. For nodes, similarity is typically based on dynamic interaction patterns or induced time-aware embeddings. For time slices or edges, similarity may involve temporal covariation or higher-order group structure.

A standard two-step node-TGC procedure is: (a) compute temporal node embeddings via a temporal graph encoder $f_\theta$ (which processes each node’s or edge’s event history), then (b) apply K-means or a similar clustering method to these embeddings, optimizing

$$\min_{c(\cdot)} \sum_{v\in V} \|h_v - \mu_{c(v)}\|^2$$

with $h_v = f_\theta(\mathrm{history}(v))$ and cluster centroid $\mu_j$ (Liu et al., 2023, Liu et al., 2023, Liu et al., 19 Jan 2026).

Clusters may be required to exhibit temporal persistence, temporal smoothness, or track emergence/disappearance. Extended problem settings include edge-clustering for hypergraphs (DeWolfe et al., 3 Jun 2025), temporal ordered clustering (with an explicit partial order constraint) (Turowski et al., 2019), and per-time-slice cluster flow tracking for topic or regime evolution (Jing et al., 2021, Nji et al., 16 Sep 2025).

2. Representative Methodological Approaches

2.1 Spectral and Laplacian-based Methods

Spatio-temporal spectral clustering uses generalizations of classical Laplacian approaches. The spatio-temporal graph Laplacian $L_{ST}$, introduced via a multiview canonical correlation analysis (CCA) framework, encodes both spatial and temporal coherence. Given a sequence of adjacency matrices $W_t$ ($t=1,\ldots,M$), one constructs block matrices $A$, $B$ and considers the generalized eigenproblem $A\mathbf{f} = \lambda B \mathbf{f}$, leading to a block Laplacian whose eigenvectors capture clusters persisting or evolving over time. Notably, for $L_{ST}=I_{nM}-C$ ($C=B^{-1}A$), eigenvalues are symmetric about 1 and indicate time-persistent cluster structure (Trower et al., 2024).

Extensions enforce temporal label smoothness, penalizing rapid label fluctuation across time by regularization (e.g., $l_1$/total variation terms on label changes) (Fukumoto et al., 2023). Primal-dual splitting or block-proximal algorithms efficiently solve the resulting constrained convex objectives.

2.2 Probabilistic and Statistical Models

Model-based TGC leverages mixture models over time-evolving graph distributions, most notably temporal exponential-family random graph models (TERGMs). Here, the clustering corresponds to latent block assignments $z_i$, while the temporal graph generation is captured by Markovian dependencies on prior snapshots and block-specific parameters. A variational EM algorithm with a mean-field approximation yields scalable inference, and model-selection via conditional likelihood BIC allows for principled determination of $K$ (Lee et al., 2017).

Ordered TGC seeks clusterings compatible with partial (or total) temporal orders, maximizing the precision at fixed density over pairs, and employs integer programming or sequential importance sampling of plausible node orderings (Turowski et al., 2019).

2.3 Matrix Factorization and Graph-of-Graphs

An alternative view is to treat the time series of graphs as a collection requiring clustering. After temporal (time windowing) and spatial discretization (vertex contraction), singular value thresholding denoises per-slice adjacency matrices; nonnegative matrix factorization (NMF) decomposes the sequence into prototypical 'motifs' and associated time-varying weights. Model selection proceeds via an (adjusted) AIC criterion (Lee et al., 2014).

2.4 Deep and Hybrid Approaches

Deep temporal graph clustering frameworks integrate temporal encoders (e.g., Hawkes process based, GNNs, or temporal attention models) with deep clustering modules (DEC-style, reconstruction, contrastive, batch-level constraints). Optimization is batch-wise over time-ordered events, accommodating memory- and speed-efficient processing (complexity $O(|E|)$ per epoch), and architectural innovations enable scaling to graphs with $>10^6$ nodes and edges (Liu et al., 2023, Liu et al., 19 Jan 2026).

Recent benchmarks such as BenchTGC (Liu et al., 19 Jan 2026) systematically retrofit static deep clustering heads with temporal GNNs, introducing feature-reconstruction, distribution alignment, cross-batch tracking, and scale regularization losses, supporting robust, time-consistent clustering without hand-tuned interlayer parameters or hand tuning for cluster number.

Ensemble and hybrid models (e.g., HEDGTC) combine homogeneous and heterogeneous clustering algorithms using dual consensus (co-occurrence + NMF), followed by a deep graph attention autoencoder (GATv2+LSTM) to further denoise and integrate spatiotemporal dependencies before latent-space clustering (Nji et al., 2024). Transformer-based models (B-TGAT) integrate U-Net and bi-directional graph attention to address high-dimensional spatiotemporal clustering with interpretable latent spaces (Nji et al., 16 Sep 2025).

3. Core Challenges and Computational Considerations

TGC exhibits unique challenges beyond those of static graph clustering:

• Knowledge leakage: Temporal models must enforce causal locality, avoiding propagation of future interactions during history encoding, which is violated by naive static GNN aggregation (Liu et al., 19 Jan 2026).
• Batch processing and time-space tradeoff: Processing sequential batches enables time-space scaling, with memory usage $O(\mathrm{batch\,size}\cdot d)$, contrasted with $O(N^2)$ for adjacency-matrix-reliant static approaches (Liu et al., 2023, Liu et al., 19 Jan 2026). This permits operation on very large graphs by trading time for space via batch size adjustment.
• Temporal persistence and smoothness: Clustering objectives frequently incorporate regularization on temporal variation, penalizing label switching, or explicitly targeting detection of evolving or splitting communities (Trower et al., 2024, Fukumoto et al., 2023).
• Model and parameter selection: Determining $K$ remains challenging; information-theoretic (AIC/BIC) and elbow-method procedures are used, but fully adaptive solutions (unknown $K$) are an open problem (Lee et al., 2014, Liu et al., 19 Jan 2026).
• Overlapping clusters and higher-order interactions: Edge-centric and hypergraph models allow flexible, overlapping community assignment and capture bursty, higher-order group events (DeWolfe et al., 3 Jun 2025).
• Data sparsity, label scarcity, and class imbalance: Datasets with unreliable or binary labels do not admit meaningful multiway clustering evaluation; large-scale, reliably labeled datasets (e.g., arXiv4TGC) have recently been introduced to resolve this bottleneck (Liu et al., 2023).

4. Benchmark Datasets, Evaluation Protocols, and Empirical Findings

Recent work has addressed limitations of prior small- or ill-annotated datasets by curating labeled, large-scale temporal graph datasets. The arXiv4TGC (AI, CS, Math, Physics, Large) datasets encompass up to $1.3$M nodes and $13$M temporal interactions with 5–172 classes, constructed via subsetting and relabeling of OGBN-Papers100M with timestamped citation events (Liu et al., 2023).

For evaluation, standard metrics include external indices (NMI, ARI, F1, ACC; ground-truth required) and unsupervised/internal indices (Calinski-Harabasz, Davies-Bouldin, silhouette, inter-cluster separation). Experimental evidence establishes that temporal models (e.g., TGC, BenchTGC) reliably outperform static clustering and temporal baselines on both accuracy and scalability, especially on data with high class overlap and size. Static methods fail (OOM or long runtime) on datasets with $>10^5$ nodes; temporal approaches remain within feasible GPU footprints (Liu et al., 19 Jan 2026).

Results from (Liu et al., 19 Jan 2026, Liu et al., 2023, Liu et al., 2023) establish that properly designed temporal models (HTNE, TREND, TGAT, TGN) with deep clustering heads achieve relative ARI and NMI improvements as high as 42% versus static or non-temporalized baselines. On dense interaction graphs (e.g., School), only temporal models achieve perfect clustering by all indices.

Edge-centric and hypergraph temporal clustering identifies functionally interpretable, temporally persistent, and overlapping communities at scale, as validated by author–paper datasets (1.28M hyperedges) (DeWolfe et al., 3 Jun 2025).

5. Domain-Specific Extensions and Specialized Applications

5.1 Topic and Regime Evolution

In social media and text-mining, TGC is used to track topic transitions by constructing temporal graphs of words per time window, clustering via Markov clustering with node-removal for modularity maximization, and quantifying cluster transitions (merge, split, unchanged, re-emerge) over time (Jing et al., 2021).

5.2 Multivariate Spatiotemporal Data

For high-dimensional climate and environmental datasets, hybrid architectures integrate spatial graph construction, ConvLSTM-based spatiotemporal encoding, and transformer or attention-driven temporal aggregation. These yield temporally stable, physically interpretable clusters corresponding to known climate regimes or dynamical transitions, with strong performance on internal structure metrics (Nji et al., 2024, Nji et al., 16 Sep 2025).

5.3 Privacy-Aware and Federated Learning

Federated temporal graph clustering (FTGC) addresses privacy constraints by decentralizing GNN-based encoder training. Clients compute local temporal embeddings and clustering objectives, with updates aggregated globally via FedAvg. Communication is optimized via quantization and sparsification. Empirical evaluations confirm that FTGC matches or exceeds centralized clustering quality without raw data exchange (Zhou et al., 2024).

6. Open Problems, Limitations, and Future Directions

Outstanding issues and future avenues include:

• Online/streaming TGC: Incremental clustering for graphs with unbounded or continuous event streams, without access to a global interaction history, remains a challenging problem (Liu et al., 19 Jan 2026).
• Cluster number adaptivity: Eliminating dependence on pre-specified $K$ and enabling one-shot or fully unsupervised cluster-number discovery.
• Overlapping and mixed-membership modeling: Realistic community structures often overlap; extending TGC to non-disjoint cluster assignments and time-varying mixed membership is necessary.
• Robustness to imbalance and sparsity: Large real datasets exhibit highly variable class frequencies and sparsity; new methods for class-balanced losses and scalable, robust clustering are being developed (Liu et al., 2023).
• Interpretable, physics-guided, or domain-specific constraints: Incorporating outside knowledge (e.g., in climate, biology, or record linkage) via plausibility functions, physics-informed constraints, or causal models remains an active direction (Nanayakkara et al., 2018, Nji et al., 16 Sep 2025).

The rapid evolution of TGC methods, benchmarking resources, and deployment in practical settings (e.g., BenchTGC, arXiv4TGC) promises further advances in dynamic network modeling, anomaly detection, recommendation, and functional segmentation of biological/sensor or textual data (Liu et al., 19 Jan 2026, Liu et al., 2023).

• "Clustering Time-Evolving Networks Using the Spatio-Temporal Graph Laplacian" (Trower et al., 2024)
• "Deep Temporal Graph Clustering" (Liu et al., 2023)
• "Model-Based Clustering of Time-Evolving Networks through Temporal Exponential-Family Random Graph Models" (Lee et al., 2017)
• "Deep Temporal Graph Clustering: A Comprehensive Benchmark and Datasets" (Liu et al., 19 Jan 2026)
• "arXiv4TGC: Large-Scale Datasets for Temporal Graph Clustering" (Liu et al., 2023)
• "Techniques for clustering interaction data as a collection of graphs" (Lee et al., 2014)
• "Hybrid Ensemble Deep Graph Temporal Clustering for Spatiotemporal Data" (Nji et al., 2024)
• "Detecting Patterns of Interaction in Temporal Hypergraphs via Edge Clustering" (DeWolfe et al., 3 Jun 2025)
• "Temporal Ordered Clustering in Dynamic Networks: Unsupervised and Semi-supervised Learning Algorithms" (Turowski et al., 2019)
• "Federated Temporal Graph Clustering" (Zhou et al., 2024)
• "Clustering of Time-Varying Graphs Based on Temporal Label Smoothness" (Fukumoto et al., 2023)
• "Temporal graph-based clustering for historical record linkage" (Nanayakkara et al., 2018)
• "B-TGAT: A Bi-directional Temporal Graph Attention Transformer for Clustering Multivariate Spatiotemporal Data" (Nji et al., 16 Sep 2025)
• "Tracing Topic Transitions with Temporal Graph Clusters" (Jing et al., 2021)