TESS Framework Improves Dynamic Graph Link Prediction Across Multiple Real-World Benchmarks

Researchers have developed TESS, short for Temporal-Embedding-Structural Synergy, a dynamic link prediction framework that fuses temporal embeddings with structural features. In network analysis, predicting links within dynamic graphs represents a fundamental task. While Graph Neural Networks have demonstrated superior performance in static link prediction settings, most established methods struggle to effectively capture the time-evolving topologies and interaction patterns inherent to dynamic networks.

Existing dynamic modeling techniques typically suffer from notable structural and temporal limitations. Current approaches frequently overlook temporal dependencies altogether or treat timing information merely as auxiliary features, failing to establish an efficient synergy between global temporal patterns and local structural characteristics.

To overcome these shortcomings, TESS generates time-sensitive node embeddings using temporal-aware random walks. The architecture employs a gated fusion mechanism to adaptively merge global temporal information with local structural features, while a dedicated structural compensation module rectifies topological information loss that occurs during the fusion process.

Experimental evaluations across nine real-world dynamic datasets demonstrated that TESS attained the highest average Area Under the Curve and ranked first on seven of the nine datasets. These results held against a broad collection of static and dynamic link prediction baselines, including recent continuous-time models, while ablation studies confirmed the necessity and effectiveness of each individual module within the framework.

Subject of Research: Dynamic graph link prediction in network analysis

Article Title: TESS: dynamic graph link prediction based on temporal awareness and feature compensation fusion

Article References: Chang, Y., Tao, Z., Wang, M., Cui, J., Yu, Q., & Tang, K. (2026). TESS: dynamic graph link prediction based on temporal awareness and feature compensation fusion. International Journal of Machine Learning and Cybernetics, 17(11), Article 524. https://doi.org/10.1007/s13042-026-03361-5

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03361-5

Keywords: dynamic graph link prediction, graph neural networks, temporal embeddings, structural compensation, gated fusion mechanism, temporal-aware random walks, continuous-time models, network analysis, TESS, dynamic, graph, link