Researchers have introduced DyGnROLE, a novel Transformer-based architecture designed for edge classification on directed dynamic graphs. This model addresses the limitations of existing architectures by disentangling source and destination node representations, capturing distinct structural and temporal contexts through separate embedding tables and role-semantic positional encodings. A key innovation is the Directional Role Alignment (DRA) self-supervised pretraining objective, which learns aligned source and destination embedding spaces, particularly effective in settings with limited labeled data. Evaluations across four edge classification tasks and eight datasets show DyGnROLE consistently outperforms state-of-the-art baselines, underscoring the benefits of role-aware learning and asymmetric pretraining for complex directed interactions. AI
IMPACT Introduces a new architecture and pretraining method for improved edge classification on dynamic graphs, particularly in low-label scenarios.
RANK_REASON The cluster contains a research paper detailing a new model architecture and pretraining objective for graph neural networks. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Directional Role Alignment
- DyGnROLE
- Gotit.pub
- Hugging Face
- ScienceCast
- Tyler Bonnett
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
