Graph Neural Networks With Parallel Neighborhood Aggregations for Graph Classification

被引:5
作者
Doshi, Siddhant [1 ]
Chepuri, Sundeep Prabhakar [1 ]
机构
[1] Indian Inst Sci, Dept Elect Commun Engn, Bangalore 560012, Karnataka, India
关键词
Computational modeling; Task analysis; Numerical models; Training; Predictive models; Computer architecture; Brain modeling; Graph classification; graph filterbanks; graph neural networks; isomorphism test; representation learning;
D O I
10.1109/TSP.2022.3205476
中图分类号
TM [电工技术]; TN [电子技术、通信技术];
学科分类号
0808 ; 0809 ;
摘要
We focus on graph classification using a graph neural network (GNN) model that precomputes node features using a bank of neighborhood aggregation graph operators arranged in parallel. These GNN models have a natural advantage of reduced training and inference time due to the precomputations but are also fundamentally different from popular GNN variants that update node features through a sequential neighborhood aggregation procedure during training. We provide theoretical conditions under which a generic GNN model with parallel neighborhood aggregations (PA-GNNs, in short) are provably as powerful as the well-known Weisfeiler-Lehman (WL) graph isomorphism test in discriminating non-isomorphic graphs. Although PA-GNN models do not have an apparent relationship with the WL test, we show that the graph embeddings obtained from these two methods are injectively related. We then propose a specialized PA-GNN model, called simple and parallel graph isomorphism network (SPIN), which obeys the developed conditions. We demonstrate via numerical experiments that the developed model achieves state-of-the-art performance on many diverse real-world datasets while maintaining the discriminative power of the WL test and the computational advantage of preprocessing graphs before the training process.
引用
收藏
页码:4883 / 4896
页数:14
相关论文
共 46 条
[1]  
[Anonymous], 2017, P INT C LEARN REPR T
[2]  
Babai L., 1979, 20th Annual Symposium of Foundations of Computer Science, P39, DOI 10.1109/SFCS.1979.8
[3]   Protein function prediction via graph kernels [J].
Borgwardt, KM ;
Ong, CS ;
Schönauer, S ;
Vishwanathan, SVN ;
Smola, AJ ;
Kriegel, HP .
BIOINFORMATICS, 2005, 21 :I47-I56
[4]   Geometric Deep Learning Going beyond Euclidean data [J].
Bronstein, Michael M. ;
Bruna, Joan ;
LeCun, Yann ;
Szlam, Arthur ;
Vandergheynst, Pierre .
IEEE SIGNAL PROCESSING MAGAZINE, 2017, 34 (04) :18-42
[5]  
Chen L., 2021, PROC INT C LEARN REP
[6]   On the Lambert W function [J].
Corless, RM ;
Gonnet, GH ;
Hare, DEG ;
Jeffrey, DJ ;
Knuth, DE .
ADVANCES IN COMPUTATIONAL MATHEMATICS, 1996, 5 (04) :329-359
[7]  
Dai HJ, 2017, ADV NEUR IN, V30
[8]  
De Cao N, 2018, PROC ICML WORKSHOP T
[9]   Distinguishing enzyme structures from non-enzymes without alignments [J].
Dobson, PD ;
Doig, AJ .
JOURNAL OF MOLECULAR BIOLOGY, 2003, 330 (04) :771-783
[10]   Graph Signal Processing for Machine Learning: A Review and New Perspectives [J].
Dong, Xiaowen ;
Thanou, Dorina ;
Toni, Laura ;
Bronstein, Michael ;
Frossard, Pascal .
IEEE SIGNAL PROCESSING MAGAZINE, 2020, 37 (06) :117-127