Graph2Video: Leveraging Video Models to Model Dynamic Graph Evolution

Hua Liu; Yanbin Wei; Fei Xing; Tyler Derr; Haoyu Han; Yu Zhang

cs.CVMar 9, 2026

Graph2Video: Leveraging Video Models to Model Dynamic Graph Evolution

Authors: Hua Liu, Yanbin Wei, Fei Xing, Tyler Derr, Haoyu Han, Yu Zhang

Organizations: Southern University of Science and Technology · The Hong Kong University of Science and Technology · City University of Hong Kong · Vanderbilt University ·

Abstract

Dynamic graphs are common in real-world systems such as social media, recommender systems, and traffic networks. Existing dynamic graph models for link prediction often fall short in capturing the complexity of temporal evolution. They tend to overlook fine-grained variations in temporal interaction order, struggle with dependencies that span long time horizons, and offer limited capability to model pair-specific relational dynamics. To address these challenges, we propose Graph2Video, a video-inspired framework that views the temporal neighborhood of a target link as a sequence of "graph frames". By stacking temporally ordered subgraph frames into a "graph video", Graph2Video leverages the inductive biases of video foundation models to capture both fine-grained local variations and long-range temporal dynamics. It generates a link-level embedding that serves as a lightweight and plug-and-play link-centric memory unit. This embedding integrates seamlessly into existing dynamic graph encoders, effectively addressing the limitations of prior approaches. Extensive experiments on benchmark datasets show that Graph2Video outperforms state-of-the-art baselines on the link prediction task in most cases. The results highlight the potential of borrowing spatio-temporal modeling techniques from computer vision as a promising and effective approach for advancing dynamic graph learning.

Figures & tables

NSS	Datasets	TGN	CAWN	TCL	GraphMixer	DyGFormer	DyGFormer+	TPNet	TPNet+
	Reddit	$98.63_{\pm 0.06}$	$99.11_{\pm 0.01}$	$97.53_{\pm 0.02}$	$97.31_{\pm 0.01}$	$99.22_{\pm 0.01}$	$99.23_{\pm 0.01}$	$\underline{99.27_{\pm 0.00}}$	99.32 ±0.01
	MOOC	$89.15_{\pm 1.60}$	$80.15_{\pm 0.25}$	$82.38_{\pm 0.24}$	$82.78_{\pm 0.15}$	$87.52_{\pm 0.49}$	$87.33_{\pm 0.43}$	96.39 ±0.09	$96.38_{\pm 0.03}$
rnd	Enron	$86.53_{\pm 1.11}$	$89.56_{\pm 0.09}$	$79.70_{\pm 0.71}$	$82.25_{\pm 1.02}$	$92.47_{\pm 0.12}$	$92.32_{\pm 0.18}$	$92.90_{\pm 0.17}$	93.01 ±0.10
	UCI	$92.34_{\pm 1.04}$	$95.18_{\pm 0.06}$	$89.57_{\pm 1.63}$	$93.25_{\pm 0.57}$	$95.79_{\pm 0.17}$	$96.24_{\pm 0.04}$	$97.35_{\pm 0.04}$	97.37 ±0.10
	Can. Parl.	$70.28_{\pm 2.34}$	$69.82_{\pm 2.34}$	$68.67_{\pm 2.67}$	$77.04_{\pm 0.96}$	$97.36_{\pm 0.45}$	$97.60_{\pm 0.40}$	$90.28_{\pm 0.37}$	98.70 ±0.13
	Avg. rank	5.60	6.00	7.60	6.40	3.60	3.40	2.20	1.20

Table 1: AP (%) for transductive dynamic link prediction on real-world datasets with three sampling strategies (NSS).

NSS	Datasets	TGN	CAWN	TCL	GraphMixer	DyGFormer	DyGFormer+	TPNet	TPNet+
	Reddit	$97.50_{\pm 0.07}$	$98.62_{\pm 0.01}$	$94.09_{\pm 0.07}$	$95.26_{\pm 0.02}$	$98.84_{\pm 0.02}$	$98.86_{\pm 0.02}$	$98.86_{\pm 0.01}$	98.95 ±0.03
	MOOC	$89.04_{\pm 1.17}$	$81.42_{\pm 0.24}$	$82.38_{\pm 0.24}$	$81.41_{\pm 0.21}$	$86.96_{\pm 0.43}$	$86.92_{\pm 0.28}$	$95.07_{\pm 0.26}$	95.10 ±0.10
rnd	Enron	$77.94_{\pm 1.02}$	$86.35_{\pm 0.51}$	$79.70_{\pm 0.71}$	$75.88_{\pm 0.48}$	$89.76_{\pm 0.34}$	$89.88_{\pm 0.12}$	90.34 ±0.28	$90.18_{\pm 0.27}$
	UCI	$88.12_{\pm 2.05}$	$92.73_{\pm 0.06}$	$89.57_{\pm 1.63}$	$91.19_{\pm 0.42}$	$94.54_{\pm 0.12}$	$94.89_{\pm 0.04}$	$95.74_{\pm 0.05}$	95.83 ±0.16
	Can. Parl.	$70.88_{\pm 2.34}$	$69.82_{\pm 2.34}$	$68.67_{\pm 2.67}$	$77.04_{\pm 0.46}$	$87.74_{\pm 0.71}$	$87.89_{\pm 1.01}$	$68.09_{\pm 1.55}$	95.50 ±0.49
	Avg. rank	5.80	5.60	6.80	6.60	3.80	3.00	3.00	1.20

Table 2: AP (%) for inductive dynamic link prediction on real-world datasets with three sampling strategies (NSS).

Figure 2: Graph Visualizer and Video Encoder Comparison.

Model	s	Enron	UCI
DyGFormer	8	92.35 ±0.20	96.19 ±0.14
	16	92.32 ±0.18	96.24 ±0.04
	32	92.31 ±0.16	96.10 ±0.38
TPNet	8	93.05 ±0.18	97.23 ±0.21
	16	93.01 ±0.10	97.37 ±0.10
	32	93.02 ±0.19	97.30 ±0.10

Table 3: Performance comparison when varying the number of most recent neighbors

s

in the video frame.

Model	Fusion Strategy	Enron	UCI
DyGFormer	MLP	92.40 ±0.12	96.23 ±0.06
	Bilinear	91.60 ±0.29	96.15 ±0.10
	Attention	92.32 ±0.18	96.24 ±0.04
TPNet	MLP	92.98 ±0.08	97.14 ±0.24
	Bilinear	93.18 ±0.12	97.04 ±0.21
	Attention	93.01 ±0.10	97.37 ±0.10

Table 4: Fusion strategies performance comparison.

Model	Sam2-F	Enron	UCI
DyGFormer	8	92.22 ±0.13	95.93 ±0.06
	16	92.25 ±0.13	95.98 ±0.08
	32	92.31 ±0.14	96.01 ±0.09
TPNet	8	92.82 ±0.09	97.16 ±0.17
	16	92.85 ±0.23	97.19 ±0.29
	32	93.01 ±0.11	97.29 ±0.12

Table 5: Performance comparison when varying the number of Frames

F

in the Sam2 video model.

Datasets	Domains	#Nodes	#Links	#N&L Feat	Bipartite	Duration	Unique Steps	Time Granularity
Reddit	Social	10,984	672,447	– & 172	True	1 month	669,065	Unix timestamps
MOOC	Interaction	7,144	411,749	– & 4	True	17 month	345,600	Unix timestamps
Enron	Social	184	125,235	– & –	False	3 years	22,632	Unix timestamps
UCI	Social	1,899	59,835	– & –	False	196 days	58,911	Unix timestamps
Can. Parl.	Politics	734	74,478	– & 1	False	14 years	14	years

Table 6: Statistics of the five real-world datasets.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

NSS	Datasets	TGN	CAWN	TCL	GraphMixer	DyGFormer	DyGFormer+	TPNet	TPNet+
	Reddit	$98.60_{\pm 0.06}$	$99.01_{\pm 0.01}$	$97.42_{\pm 0.02}$	$97.17_{\pm 0.02}$	$99.15_{\pm 0.01}$	$99.16_{\pm 0.01}$	$99.22_{\pm 0.00}$	$\textbf{99.27}_{\pm 0.01}$
	MOOC	$91.21_{\pm 1.15}$	$80.38_{\pm 0.26}$	$83.12_{\pm 0.18}$	$84.01_{\pm 0.17}$	$87.91_{\pm 0.58}$	$86.93_{\pm 0.38}$	$\textbf{97.17}_{\pm 0.08}$	$97.13_{\pm 0.02}$
rnd	Enron	$88.32_{\pm 0.99}$	$90.45_{\pm 0.14}$	$75.74_{\pm 0.72}$	$84.38_{\pm 0.21}$	$93.33_{\pm 0.13}$	$93.45_{\pm 0.13}$	$93.98_{\pm 0.26}$	$\textbf{94.16}_{\pm 0.10}$
	UCI	$92.03_{\pm 1.13}$	$93.87_{\pm 0.08}$	$87.82_{\pm 1.36}$	$91.81_{\pm 0.67}$	$94.49_{\pm 0.26}$	$95.22_{\pm 0.07}$	$96.79_{\pm 0.05}$	$\textbf{96.86}_{\pm 0.14}$
	Can. Parl.	$76.99_{\pm 1.80}$	$75.70_{\pm 3.27}$	$72.46_{\pm 3.23}$	$83.17_{\pm 0.53}$	$97.76_{\pm 0.41}$	$97.99_{\pm 0.41}$	$92.05_{\pm 0.34}$	$\textbf{98.83}_{\pm 0.12}$
	Avg. rank	5.40	6.00	7.60	6.60	3.80	3.20	2.20	1.20

Appendix

Table 7: AUC (%) for transductive dynamic link prediction on real-world datasets with three sampling strategies (NSS).

NSS	Datasets	TGN	CAWN	TCL	GraphMixer	DyGFormer	DyGFormer+	TPNet	TPNet+
	Reddit	$97.39_{\pm 0.07}$	$98.42_{\pm 0.02}$	$93.80_{\pm 0.07}$	$94.97_{\pm 0.05}$	$98.71_{\pm 0.01}$	$98.71_{\pm 0.01}$	$\underline{98.73_{\pm 0.02}}$	$\textbf{98.83}_{\pm 0.03}$
	MOOC	$91.24_{\pm 0.99}$	$81.86_{\pm 0.25}$	$81.43_{\pm 0.19}$	$82.77_{\pm 0.24}$	$87.62_{\pm 0.51}$	$87.91_{\pm 0.24}$	$\textbf{95.55}_{\pm 0.25}$	$\textbf{95.55}_{\pm 0.15}$
rnd	Enron	$78.83_{\pm 1.11}$	$87.02_{\pm 0.50}$	$72.33_{\pm 0.99}$	$76.51_{\pm 0.71}$	$\underline{90.69_{\pm 0.26}}$	$\textbf{90.78}_{\pm 0.17}$	$90.21_{\pm 0.49}$	$90.18_{\pm 0.48}$
	UCI	$86.68_{\pm 2.29}$	$90.40_{\pm 0.11}$	$84.49_{\pm 1.82}$	$89.30_{\pm 0.57}$	$92.63_{\pm 0.13}$	$93.27_{\pm 0.03}$	$\underline{94.40_{\pm 0.03}}$	$\textbf{94.52}_{\pm 0.26}$
	Can. Parl.	$55.86_{\pm 0.75}$	$58.83_{\pm 1.13}$	$55.83_{\pm 1.07}$	$58.32_{\pm 1.08}$	$89.33_{\pm 0.48}$	$\underline{89.51_{\pm 0.82}}$	$69.21_{\pm 1.31}$	$\textbf{95.69}_{\pm 0.50}$
	Avg. rank	5.80	5.40	8.00	6.40	3.40	2.80	2.60	1.60

Appendix

Table 8: AUC (%) for inductive dynamic link prediction on real-world datasets with three sampling strategies (NSS).

Dynamic Graph Learning

Dynamic graph learning models how networks and their node features change over time to predict links, events, labels, and system behavior.

See all 124 papers

Feb 15, 2026

A Hybrid TGN-SEAL Model for Dynamic Graph Link Prediction

Predicting links in sparse, continuously evolving networks is a central challenge in network science. Conventional heuristic methods and deep learning models, including Graph Neural Networks (GNNs), are typically designed for static graphs and thus struggle to capture temporal dependencies. Snapshot-based techniques partially address this issue but often encounter data sparsity and class imbalance, particularly in networks with transient interactions such as telecommunication call detail records (CDRs). Temporal Graph Networks (TGNs) model dynamic graphs by updating node embeddings over time; however, their predictive accuracy under sparse conditions remains limited. In this study, we improve the TGN framework by extracting enclosing subgraphs around candidate links, enabling the model to jointly learn structural and temporal information. Experiments on a sparse CDR, email, message dataset show that our approach increases average precision by at least 2% over standard TGNs, demonstrating the advantages of integrating local topology for robust link prediction in dynamic networks.

Nafiseh Sadat Sajadi, Behnam Bahrak, Mahdi Jafari Siavoshani

Department of Computer Engineering, Sharif University of Technology, Tehran, Iran. · Tehran Institute for Advanced Studies, Khatam University, Tehran, Iran.

Temporal GNNs Link Prediction Temporal Link Prediction Dynamic Graph Learning

Ask PDF

Oct 14, 2026

Dynamic Graph World Modelling: Learning Future Dynamics from Temporal Context

Dynamic graph learning models how entities and their relations evolve from historical observations to predict future graph states. Existing methods do not fully address dynamic graph world modelling, which requires learning transitions from temporal context to future latent graph states. This is difficult for three reasons: nodes and relations rely on different historical evidence, structural evolution is path-dependent, and future snapshots provide sparse and realization-specific targets. We propose Dynamic Graph World Modelling (DyG-WM), the first framework to jointly predict future node and relation representations from temporal context by answering three questions: what has evolved, how it evolved, and what will evolve next. To capture what has evolved, a historical node-relation context encoder summarizes node-neighborhood trajectories and pair-specific histories. To characterize how it evolved, a node-relation path-signature encoder captures ordered structural changes. To predict what will evolve next, a dual-level joint-embedding objective predicts target-snapshot node and relation representations without reconstructing the raw graph. Experiments on 12 link-prediction and two node-classification datasets demonstrate the effectiveness of DyG-WM across both tasks.

Bo Li, Yue Yang, Yujin Huang +6

School of Information and Communication Technology, Griffith University, Australia · Maincode, Australia · School of Computing and Information Systems, University of Melbourne, Australia +2

Dynamic Graph Learning World Model Learning Predictive Representation Learning Temporal Link Prediction Joint-Embedding Predictive Architecture

Ask PDF

May 28, 2026

Forget Less, Generalize More: Unifying Temporal and Structural Adaptation for Dynamic Graphs

Representation learning on dynamic graphs requires capturing complex dependencies that evolve across both time and structure. Existing approaches typically adopt fixed temporal decay schemes or predetermined structural propagation depths, limiting their ability to generalize across graphs with diverse interaction frequencies and topological characteristics. We propose Dual-Scale Retentive Dynamics (DSRD), a unified framework that maintains a retentive representation state encoding both temporal memory and structural context. DSRD introduces two key components: (i) a retentive state with dual-scale adaptation that jointly models temporal dynamics and structural propagation within a single recurrent formulation, and (ii) adaptive decay kernels with learnable time-sensitivity parameters that automatically balance short-term responsiveness and long-term retention based on the underlying interaction patterns. We provide theoretical analysis establishing the equivalence between event-wise parallel aggregation and efficient recurrent state updates, as well as stability and boundedness guarantees for the learned dynamics. Extensive experiments on 14 real-world benchmarks demonstrate that DSRD consistently achieves state-of-the-art performance on both link prediction and node classification tasks, with strong generalization across transductive and inductive settings.

Qian Chang, Ciprian Doru Giurcaneanu, Runsong Jia +6

University of Auckland, Auckland, New Zealand · University of Technology Sydney, Sydney, Australia · Central China Normal University, Wuhan, China

Temporal GNNs Graph Representation Learning Temporal Link Prediction Dynamic Graph Learning Adaptive Learning Node Classification

Ask PDF

Abstract

Dynamic Graph Learning

Explore similar work

A Hybrid TGN-SEAL Model for Dynamic Graph Link Prediction

Dynamic Graph World Modelling: Learning Future Dynamics from Temporal Context

Forget Less, Generalize More: Unifying Temporal and Structural Adaptation for Dynamic Graphs