Enhancing the Expressivity of Temporal Graph Networks through Source-Target Identification

The project addresses a specific technical gap in Temporal Graph Networks (TGNs): their inability to outperform simple moving averages in predicting the *weight* or *intensity* of future interactions (affinity) rather than just the existence of an edge. While the paper provides a meaningful theoretical insight (Source-Target Identification), the repository currently functions as a reproducibility package for the paper rather than a maintained software tool. With 0 stars and 3 forks, it has no community traction or ecosystem lock-in. Its value is purely academic/algorithmic. Frontier labs are unlikely to compete directly as this is a niche optimization for specific graph tasks, but the technique itself is easily reproducible and likely to be superseded by more general-purpose Graph Transformers or State Space Models (SSMs) adapted for graphs within 12-24 months. The low defensibility reflects its status as a research artifact rather than a platform or product.