Fast Approximate Graph Neural Networks Using Hashing Function

Graph machine learning has seen significant advancements in recent years, particularly in the domain of small graphs such as molecules. These advancements are attributed to the application of transformer architectures, which have gained widespread recognition for achieving state-of-the-art performance in various deep learning domains. These graph networks employ variations of global attention mechanisms to enable interactions between all nodes, but suffer from limited scalability. In our work, we address this issue by exploring probabilistic data structures, specifically the bloom graph, which compresses information about a set of nodes into a compact vector, thereby reducing memory consumption and improving the speed of the global attention mechanism. We anticipate that our proposed approach, which we call the Hashformer layer, will yield comparable results in terms of quality and superior performance in terms of speed.