Never Skip a Batch: Continuous Training of Temporal GNNs via Adaptive Pseudo-Supervision

The project serves as a research reference implementation for a specific optimization in Temporal Graph Neural Networks (TGNNs). While the theoretical foundation—reducing gradient variance via History-Averaged Labels (HAL)—is sound for researchers in the niche, the project shows zero organic adoption (0 stars) and minimal engagement (5 forks) nearly a year after its inception. Defensibility is nearly non-existent as the core contribution is an algorithmic technique that could be easily ported to dominant graph frameworks like PyTorch Geometric (PyG) or Deep Graph Library (DGL). Frontier labs are unlikely to compete directly as they focus on large-scale LLM architectures, leaving TGNNs to specialized academic or industrial labs (e.g., Pinterest, Uber, or Amazon's graph teams). The primary risk is obsolescence: as a standalone research repo, it will likely be superseded by the next incremental paper or integrated as a 'feature' in a more popular library, effectively ending its life as a standalone entity.