This website requires JavaScript.

Efficient Graph Neural Network Inference at Large Scale

Xinyi GaoWentao ZhangYingxia ShaoQuoc Viet Hung NguyenBin CuiHongzhi Yin
Nov 2022
摘要
Graph neural networks (GNNs) have demonstrated excellent performance in awide range of applications. However, the enormous size of large-scale graphshinders their applications under real-time inference scenarios. Althoughexisting scalable GNNs leverage linear propagation to preprocess the featuresand accelerate the training and inference procedure, these methods still sufferfrom scalability issues when making inferences on unseen nodes, as the featurepreprocessing requires the graph is known and fixed. To speed up the inferencein the inductive setting, we propose a novel adaptive propagation orderapproach that generates the personalized propagation order for each node basedon its topological information. This could successfully avoid the redundantcomputation of feature propagation. Moreover, the trade-off between accuracyand inference latency can be flexibly controlled by simple hyper-parameters tomatch different latency constraints of application scenarios. To compensate forthe potential inference accuracy loss, we further propose InceptionDistillation to exploit the multi scale reception information and improve theinference performance. Extensive experiments are conducted on four publicdatasets with different scales and characteristics, and the experimentalresults show that our proposed inference acceleration framework outperforms theSOTA graph inference acceleration baselines in terms of both accuracy andefficiency. In particular, the advantage of our proposed method is moresignificant on larger-scale datasets, and our framework achieves $75\times$inference speedup on the largest Ogbn-products dataset.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答