Transfer Learning for Edge Classification on Dynamic Text-Attributed Graphs

Tyler Bonnet, Marek Rei

arXiv: 2609.32849Published: September 26, 2026

AI summary

TL;DR

This paper proposes a transfer learning approach for edge classification on dynamic text-attributed graphs, introducing a Spatio-Temporal Semantic Alignment (STSA) model that outperforms existing methods. STSA integrates a spatio-temporal encoder and a Contrastive Semantic Forecasting objective. It achieves robust performance across domains.

Why It Matters

The proposed approach enables effective edge classification in dynamic graphs, which is crucial for various applications such as social network analysis and recommender systems. It also highlights the importance of transfer learning in graph representation learning. This can lead to more accurate and robust predictions in real-world scenarios.

Key Findings

State-of-the-art self-supervised methods perform poorly on unseen domains
The proposed STSA model outperforms existing methods
A simple Bag of Events model can be competitive with more complex models

Topics

graph neural networks

transfer learning

natural language processing

contrastive learning

multimodal learning

Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.