Towards Adaptive Neighborhood for Advancing Temporal Interaction Graph Modeling

The project is a standard academic reference implementation for a research paper (arXiv:2406.11891). While it addresses a legitimate technical bottleneck in Temporal Graph Networks (TGNs)—the reliance on rigid neighborhood sampling—it lacks any defensible moat beyond the research itself. With 0 stars and 8 forks, the project has negligible community adoption. It is a 'better component' for graph modeling rather than a standalone platform. Frontier labs like Google (DeepMind) or Meta (FAIR) frequently innovate in graph neural networks (GNNs) for recommendation systems and social maps; they are likely to adopt similar 'adaptive neighborhood' concepts internally, rendering this specific repo obsolete. The high displacement horizon reflects the rapid SOTA (state-of-the-art) churn in graph learning research, where a new sampling strategy is typically superseded within 12-24 months.