Transfer Learning for Edge Classification on Dynamic Text-Attributed Graphs
Researchers show dynamic graph models transfer poorly and propose STSA for cross-domain edge classification.
The paper defines a leave-one-domain-out transfer protocol for edge classification on dynamic text-attributed graphs and reports that state-of-the-art self-supervised dynamic graph methods transfer poorly. A structurally and temporally unaware Bag of Events model, using only unordered node and edge text, outperforms those methods. The proposed Spatio-Temporal Semantic Alignment model fuses time deltas and node occurrence frequencies and is trained with Contrastive Semantic Forecasting against a pretrained language-model text space, beating BoE and the evaluated baselines.
- A leave-one-domain-out protocol tests edge classification transfer on dynamic text-attributed graphs.
- Self-supervised dynamic graph methods underperform a text-only Bag of Events baseline.
- STSA fuses time deltas and node frequencies using contrastive semantic forecasting.
From the original article
1 min read at arxiv.orgLearning transferable representations for dynamic text-attributed graphs (DyTAGs) requires models to capture underlying interaction dynamics that persist across domains. However, existing methods tend to overfit to domain-specific structural, temporal, and semantic patterns, limiting edge classification performance under distribution shifts.
By Tyler Bonnet · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.