Fast and Accurate Real-time Inference with Graph-less Neural Networks

Graph Neural Networks (GNNs) are increasingly popular tools for graph machine learning tasks of recent years, including those on knowledge graphs. Yet, these models come with major scalability challenges owing to their complex data dependency, requiring multiple hops of information away from an inference target. These challenges make real-time inference with GNNs at best slow, and at worst infeasible in practical settings. To circumvent these challenges, we propose Graph-less Neural Networks -- a simple yet hugely impactful model training and inference paradigm which yields competitively accurate models which infer faster than naive GNNs by 146-273x, and faster than other acceleration methods by 14-27x across node classification benchmarks