STGraph: A Framework for Temporal Graph Neural Networks

Temporal graphs are extensively used to model interactions in domains such as e-commerce, social media, and transportation. Temporal Graph Neural Networks (TGNNs) are utilized to analyze the spatial and temporal properties of these graphs. This paper introduces STGraph, an innovative framework designed for programming TGNNs. By extending Seastar, a vertex-centric programming model for GPU-based GNN training, STGraph is capable of learning from both static graphs with temporal signals and discrete-time dynamic graphs (DTDGs). In contrast to existing TGNN frameworks, which incur substantial memory overhead by storing DTDGs as separate snapshots, STGraph dynamically constructs snapshots on demand during training. This is achieved through seamless integration with dynamic graph data structures capable of generating snapshots from temporal updates. Additionally, we present improvements to the Seastar design, for easier maintenance and greater software portability. STGraph exhibits significant performance gains when benchmarked against Pytorch Geometric Temporal (PyG-T) on an NVIDIA GPU. For static graphs with temporal signals, STGraph shows up to 1.69× speed-up and up to 2.14× memory improvement over PyG-T. For DTDGs, STGraph exhibits up to 1.20× speed-up and 1.91× memory improvement over PyG-T.