Dynamic Graph Learning Framework

• Dynamic graph learning frameworks are algorithmic systems that represent and process evolving graph structures via snapshot and continuous time formulations.
• They leverage encoder–decoder and modular neural architectures to integrate temporal dynamics with structural information for tasks like link prediction and clustering.
• They face challenges in benchmark standardization and scalability, motivating ongoing research into robust, efficient, and unified evaluation protocols.

Searching arXiv for relevant papers on dynamic graph learning frameworks and benchmarks. Search query: dynamic graph learning framework arXiv A dynamic graph learning framework is a formal and algorithmic apparatus for representing, encoding, predicting, and evaluating graphs whose nodes, edges, and attributes evolve over time. In the literature, the term is used in several senses: as a general encoder–decoder abstraction for supervised dynamic graph learning, as a modular neural architecture for continuous-time or snapshot-based graphs, and as an integrated ecosystem combining data representation, modeling, tasks, metrics, and experimental protocols (Zhu et al., 2022, Zhang, 2024). Across these uses, the central problem is consistent: to transform temporal graph observations into representations such as $\mathbf{z}_i(t)$ or $h_{v,t}$ that preserve both structure and temporal dynamics while supporting downstream objectives such as link prediction, node classification, clustering, graph generation, and anomaly detection (Rossi et al., 2020, Zhang, 2024).

1. Formalization and temporal representations

A static graph is commonly written as $G=(V,E,X)$, where $V$ is the node set, $E \subseteq V \times V$ is the edge set, and $X$ denotes node features. In a dynamic graph, the nodes, edges, and/or their attributes change over time, and the input at each discrete time step is formalized as

$G_t = (V_t, E_t, X_t),$

with the graph evolving as a sequence $\{G_t\}_{t=1}^T$ (Zhang, 2024). This snapshot-based formulation is the canonical representation for discrete-time dynamic graphs.

A second representation treats the graph as a continuous-time event stream. In this setting, a temporal graph is modeled as a sequence of time-stamped events

$$\mathcal{G} = \{x(t_1), x(t_2), \dots \}, \quad 0 \le t_1 \le t_2 \le \dots,$$

where events may be node-wise feature updates or interaction events between nodes (Rossi et al., 2020). The survey literature distinguishes these two storage models as discrete time dynamic graphs and continuous time dynamic graphs, and also distinguishes implicit-time models, which learn from sequence order, from explicit-time models, which encode timestamps or time differences directly (Zhu et al., 2022).

This bifurcation of representations is not merely notational. Snapshot models privilege temporally ordered graph states, whereas event-stream models privilege irregular interactions and fine-grained temporal resolution. A plausible implication is that the notion of a dynamic graph learning framework is best understood as representation-dependent: some frameworks are designed around $\{G_t\}_{t=1}^T$, others around $h_{v,t}$0 as a sequence of events, and many comparative difficulties arise precisely because the two settings are often evaluated under incompatible assumptions (Zhang, 2024).

2. Core architectural abstractions

A standard abstraction is the encoder–decoder architecture for supervised dynamic graph learning. The survey literature formalizes temporal evolution through a Three Stages Recurrent Temporal Learning Framework consisting of Attribute Self-Updating, Association Process, and Message Passing, with temporal pattern expressed as

$h_{v,t}$1

In this view, attributes evolve, edge structure changes in response to attributes and prior topology, and message passing updates node states on the resulting graph (Zhu et al., 2022).

A complementary abstraction is the general Dynamic Graph Neural Network paradigm. At each time step $h_{v,t}$2, the model maintains a hidden state $h_{v,t}$3 for each node and applies two operations:

$h_{v,t}$4

$h_{v,t}$5

where $h_{v,t}$6 is a possibly time-dependent aggregation function and $h_{v,t}$7 combines prior state, aggregated message, and current features (Zhang, 2024). This formulation subsumes many snapshot-based temporal GNNs.

For continuous-time graphs, Temporal Graph Networks make the modularity explicit. TGN comprises five core components: a memory module $h_{v,t}$8, message functions, a message aggregator, a memory updater, and an embedding module that converts memory, features, and temporal neighborhood information into node embeddings $h_{v,t}$9 (Rossi et al., 2020). Because TGN can recover JODIE, TGAT, and DyRep as special cases under specific module choices, it serves as an archetypal example of a framework in the strong architectural sense: not a single model, but a parameterized family of models.

These abstractions converge on a common principle. A dynamic graph learning framework must specify how temporal observations are represented, how state is propagated across time, how neighborhood information is aggregated under changing topology, and how the resulting representations are exposed to downstream tasks. The benchmark literature extends this principle further by arguing that the framework should also standardize preprocessing, temporal splits, negative sampling, hyperparameter tuning, and reporting conventions (Zhang, 2024).

3. Representative framework families

The literature contains several distinct framework families, differentiated by temporal granularity, inductive bias, and training objective.

Family	Core mechanism	Representative frameworks
Event-driven memory models	Node memory, event messages, temporal neighborhood operators	TGN (Rossi et al., 2020), DG-Gen (Hosseini et al., 2024)
Snapshot recurrent and tensor models	Snapshot sequence, recurrent updates, or tensor convolution	DSVB (Yap et al., 2023), TGCN (Wang et al., 2024)
Continuous-time differential models	Control paths and neural differential equations on evolving graphs	GN-CDE (Qin et al., 2023)
Self-supervised and disentangled models	Temporal subgraph contrast, temporal-clips contrast, timespan-view contrast	DySubC (Jiang et al., 2021), DyTed (Zhang et al., 2022), CLDG (Xu et al., 26 May 2026)
Higher-order and heterogeneous models	Hyperedges, meta-graphs, typed relations	LFH (Zhang et al., 2023), MG-DVD (Liu et al., 2021)

Within the snapshot family, Tensor Graph Convolutional Network represents the dynamic graph as adjacency and weight tensors, and performs graph convolution through an $G=(V,E,X)$0-product, yielding a tensor-level analogue of GCN,

$G=(V,E,X)$1

with a lower-triangular, banded temporal mixing matrix $G=(V,E,X)$2 that enforces causal temporal aggregation (Wang et al., 2024). A different snapshot-based design appears in the Deep Spatiotemporal Variational Bayes framework, which couples an attention-based message-passing encoder with a spatial-aware GRU and a sequential ELBO objective to learn probabilistic dynamic graph embeddings (Yap et al., 2023).

Continuous-time formulations exhibit a different design logic. Graph Neural Controlled Differential Equation treats the time-augmented adjacency tensor as a control path and defines node trajectories through

$G=(V,E,X)$3

The paper emphasizes that this allows dynamics on evolving graphs without integration by segments and supports calibration with subsequent data as well as robustness to missing observations (Qin et al., 2023). This suggests a broader trend in dynamic graph learning: sequence models and recurrent updates are no longer the only mechanisms for temporal state evolution.

Self-supervised frameworks introduce another axis of variation. DySubC defines a temporal subgraph contrastive learning task for continuous-time graphs by sampling temporal subgraphs around each node, encoding them with GCNs, and maximizing mutual information between a node representation and its temporal subgraph representation, while also contrasting against temporal negatives derived from time-agnostic subgraphs (Jiang et al., 2021). DyTed, by contrast, explicitly separates time-invariant and time-varying information, learning $G=(V,E,X)$4 and $G=(V,E,X)$5 such that

$G=(V,E,X)$6

and then enforcing disentanglement through temporal-clips contrastive learning and an adversarial discriminator motivated by mutual-information minimization (Zhang et al., 2022). CLDG and CLDG++ adopt yet another inductive bias, temporal translation invariance, and perform contrastive learning over different timespan views; CLDG++ further adds graph diffusion and local-local, local-global, and global-global contrasts (Xu et al., 26 May 2026).

Higher-order frameworks shift focus from pairwise edges to more expressive relational units. LFH, a hypergraph learning framework, performs pairwise fusion for initial node embedding, constructs implicit hyperedges through dynamic grouping, and then applies type-specific hypergraph learning with attentive embedding update. In its reported experiments it yields average gains of 12.5% in node classification and 13.3% in link prediction over recent state-of-the-art methods (Zhang et al., 2023). MG-DVD models malware execution event streams as dynamic heterogeneous graphs, introduces real-world meta-graphs between malware objects, and employs two dynamic walk-based heterogeneous graph learning methods to reduce graph retraining cost in real time (Liu et al., 2021).

4. Specialized extensions of the framework notion

Several recent works expand the meaning of framework beyond temporal representation learning proper and treat it as a scaffold for continual learning, topology-aware adaptation, generation, or collaboration.

In dynamic graph continual learning, CCC models a dynamic graph as a sequence of snapshots $G=(V,E,X)$7 and addresses catastrophic forgetting through a condensation–concatenation pipeline. Historical snapshots are compressed into condensed graphs that preserve label distribution and topological properties, a dynamic model such as EvolveGCN is trained on the condensed sequence, and historical embeddings are selectively concatenated with current embeddings for nodes in structural change regions. CCC also introduces a node-level forgetting measure,

$G=(V,E,X)$8

to quantify predictive degradation on existing nodes caused by structural updates (Yan et al., 12 Dec 2025).

Topology-aware meta-learning appears in TMetaNet, which treats dynamic link prediction from snapshot $G=(V,E,X)$9 to $V$0 as a per-snapshot task and modulates parameter updates using differences in high-order topological signatures extracted by Dowker Zigzag Persistence. Its update rule is

$V$1

where $V$2 is produced by a CNN and fully connected layer operating on differences between zigzag persistence images of adjacent time windows (Li et al., 31 May 2025). This is notable because it turns topology evolution itself into a meta-learning signal.

Generative formulations redefine the framework around event likelihood rather than representation alone. DG-Gen works on continuous-time dynamic graphs represented as event sequences and factorizes the joint probability of an interaction as

$V$3

Time is modeled through an Exponential distribution over inter-event time, while edge features are generated autoregressively with categorical or Gaussian-mixture outputs. The resulting model is an encoder–decoder generative framework that supports both synthetic dynamic graph generation and link prediction (Hosseini et al., 2024).

Collaborative and privacy-constrained settings motivate yet another extension. DG-CoLearn works with vertex-partitioned dynamic graph snapshots under a trusted-server, client-oblivious setting. It processes only the affected $V$4-hop induced subgraph around updated nodes, reuses historical hidden states elsewhere, and performs server-mediated embedding exchange so that exact corrected 1-hop and 2-hop embeddings can be reconstructed for border nodes without exposing raw cross-client structural information. On its reported benchmarks, DG-CoLearn achieves up to 33.8$V$5 speedup in training time and 27.4$V$6 reduction in communication overhead while improving both node classification and link prediction (Au et al., 29 May 2026).

5. Tasks, metrics, and benchmarking

Dynamic graph models are typically evaluated through downstream tasks rather than intrinsic reconstruction alone. The benchmark literature enumerates temporal link prediction, dynamic node classification, node or graph clustering, graph classification, graph generation, anomaly detection, and domain-specific tasks such as information diffusion prediction and recommendation (Zhang, 2024). This task-centric view is shared across framework families: TGN is evaluated on future edge prediction and dynamic node classification (Rossi et al., 2020), DySubC on downstream link prediction (Jiang et al., 2021), CLDG on node classification and dynamic graph anomaly detection (Xu et al., 26 May 2026), and DG-Gen on graph generation fidelity and link prediction (Hosseini et al., 2024).

A central controversy is evaluation inconsistency. On the same datasets, different dynamic graph papers often report different metrics: TGN, NAT, and DGNN use AP or AUC, JODIE reports MRR, and other works report accuracy, precision, recall, F1-score, or topological metrics such as average degree and clustering coefficient (Zhang, 2024). Even Wikipedia and Reddit, which are canonical temporal interaction datasets, are evaluated under different temporal ranges, preprocessing rules, negative sampling schemes, and train/validation/test splits. The benchmark critique is therefore methodological rather than architectural: without common temporal splits, future-only evaluation, fixed dataset versions, shared negative sampling, and identical metrics, state-of-the-art comparisons are difficult to interpret (Zhang, 2024).

This benchmarking problem is not peripheral. It determines what counts as a dynamic graph learning framework in practice. The benchmark paper argues that a framework should include time-aware graph data structures, canonical dataset loaders, temporal splitting utilities, model APIs aligned with the DGNN paradigm, standardized evaluation modules for AUC, AP, precision, recall, F1, MRR, Hits@k, and topological measures, and experiment management for hyperparameter search and early stopping (Zhang, 2024). A plausible implication is that the field increasingly treats “framework” as both a modeling object and an evaluation object.

6. Domain-specific instantiations

Dynamic graph learning frameworks have been specialized for domains in which temporal relational structure is indispensable rather than optional. In multivariate time series forecasting, SDGL learns both a static adjacency $V$7 and a dynamic adjacency $V$8 from data. The static graph captures long-term association patterns via node embeddings and graph regularization, whereas the dynamic graph is built from current inputs and static embeddings through an information fusion module and a dynamic graph learning module. The model then combines TCN-based temporal modeling with GCN propagation over both $V$9 and $E \subseteq V \times V$0 (Li et al., 2021). This use case illustrates that dynamic graph learning frameworks are not confined to social or transactional networks; they also subsume spatiotemporal sensor systems.

In brain network analysis, DSVB models dynamic functional connectivity as a sequence of graphs with fixed node set and time-varying adjacency. It combines variational inference, attention-based message passing, a spatial-aware GRU, and adversarial training to learn probabilistic graph embeddings for autism spectrum disorder identification, and the reported evaluation on ABIDE shows that the framework substantially outperforms several static and dynamic baselines (Yap et al., 2023). Here the framework operates simultaneously as an unsupervised graph sequence model and a graph-level classifier.

In cybersecurity, MG-DVD constructs dynamic heterogeneous graphs from fine-grained execution event streams and introduces meta-graph-guided dynamic walks to support real-time malware variant detection. Its hybrid use of a Dynamic Walk Incremental Updating Encoder and a Combined Hierarchical Graph Attention Encoder is explicitly motivated by the need to avoid retraining the entire graph when only a local subset has changed (Liu et al., 2021). This suggests that dynamic graph learning frameworks are especially attractive when latency constraints and heterogeneous semantics coexist.

Dynamic graph anomaly detection provides another distinctive application. CLDG and CLDG++ define anomaly scores from the inconsistency of a node’s embeddings across timespan views, under the assumption of temporal translation invariance, and the paper reports desirable performance on node classification and dynamic graph anomaly detection while emphasizing reduced time and space complexity relative to complicated sequence models (Xu et al., 26 May 2026). In this domain, the framework’s inductive bias is itself reused at inference time as an anomaly criterion.

7. Limitations and future directions

Several limitations recur across otherwise dissimilar frameworks. The benchmark critique identifies the absence of a unified, standardized benchmark framework as a field-level limitation, citing inconsistent metrics, heterogeneous comparison protocols, dataset-level inconsistencies, and the lack of standardized experimental setups for temporal splitting, negative sampling, cold-start handling, and early stopping (Zhang, 2024). This limitation is methodological and remains orthogonal to improvements in modeling.

Scalability remains a technical constraint. SDGL learns dense static and dynamic adjacencies and therefore has quadratic complexity in the number of variables; the paper explicitly notes quadratic complexity in $E \subseteq V \times V$1 as a potential limitation (Li et al., 2021). TGCN reduces temporal mixing cost through a lower-triangular, banded matrix $E \subseteq V \times V$2, but full tensors $E \subseteq V \times V$3 still create memory pressure for large $E \subseteq V \times V$4 and $E \subseteq V \times V$5 (Wang et al., 2024). TMetaNet mitigates persistent-homology cost through Dowker complexes, yet still reports non-trivial overhead from computing Dowker zigzag persistence and zigzag persistence images (Li et al., 31 May 2025). DG-CoLearn addresses efficiency through incremental processing, but its privacy model assumes a trusted server and its exact cross-client message passing is restricted to 2-hop structure (Au et al., 29 May 2026).

Framework-specific inductive biases can also become liabilities. CLDG depends on temporal translation invariance, the tendency of an identical node to keep similar labels across different timespans; the paper explicitly notes that rapid label change or strong concept drift may violate this assumption (Xu et al., 26 May 2026). CCC is snapshot-based and therefore does not directly address event-level streaming without adaptation (Yan et al., 12 Dec 2025). GN-CDE’s behavior depends on the interpolation scheme used to construct the control path from irregular graph observations, which the paper identifies as a modeling choice rather than a neutral preprocessing step (Qin et al., 2023).

Current directions point toward richer but more demanding framework notions: benchmark suites that integrate data, tasks, and metrics; collaborative systems that preserve structural privacy; continual-learning systems that resist catastrophic forgetting; and topology-aware or higher-order systems that use hypergraphs, persistent homology, or diffusion views. This suggests that the term “dynamic graph learning framework” is continuing to broaden. It no longer denotes only a temporal GNN architecture; it increasingly denotes a complete research object that couples graph representation, temporal modeling, learning objective, deployment setting, and evaluation protocol into a single coherent design (Zhu et al., 2022, Zhang, 2024).