The paper studies for edge classification on dynamic text-attributed graphs (DyTAGs), where interactions between entities have text features and evolve over time. Existing self-supervised methods fail when transferred to new domains, often performing worse than a simple Bag of Events (BoE) model that ignores structure and time. The authors propose Spatio-Temporal Semantic Alignment (STSA), which fuses time deltas and node frequencies, and trains with Contrastive Semantic Forecasting to anchor representations to a multi-domain textual space from a pretrained language model. STSA consistently outperforms all baselines across multiple real-world datasets.
Semantic alignment improves edge classification across unseen dynamic graph domains
A leave-one-domain-out evaluation finds that existing dynamic-graph models transfer poorly, while aligning temporal and structural signals with pretrained textual representations improves performance.
Top University
Tyler Bonnet · Marek Rei
Imperial College London
Research Digest··2 min read
In one line
Spatio-Temporal Semantic Alignment outperforms all existing methods on edge classification for dynamic text-attributed graphs.
Before you read this
Three ideas this paper builds ontransfer learningUsing knowledge gained from one task to improve learning on a different but related task.5 papers →self-supervised learningA paradigm where models learn representations from unlabeled data by predicting hidden or transformed parts of the input.60 papers →contrastive alignmentA method that learns embeddings by minimizing distance between positive pairs and maximizing distance between negative pairs.48 papers →
Read it as
Bonnet and Rei introduce a transfer-learning protocol for classifying future interactions in dynamic text-attributed graphs from previously unseen domains.
The authors formalize a leave-one-domain-out protocol in which a model is self-supervised on several source domains, then evaluated on labeled future interactions from a held-out domain.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.