Survey on Graph Neural Network Acceleration: An Algorithmic Perspective

Nozawa, Kento; Sato, Issei

doi:10.24963/ijcai.2022/772

Cited by 16 publications

(2 citation statements)

References 2 publications

(2 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Later on, many spatial-based convolutional graph neural networks emerged [65], [66], [67]. Recently, while some researchers developed advanced algorithms such as GNN acceleration algorithms [68], some focused on real-world applications.…”

Section: Deep Learning On Graph Datamentioning

confidence: 99%

Identifying Multiple Propagation Sources With Motif-Based Graph Convolutional Networks for Social Networks

2023

View full text Add to dashboard Cite

Identifying the sources of propagation in social networks, such as the misinformation propagation, is one of the key issues recently. Most existing studies assume the underlying propagation model is known, which is difficult to obtain in practice. Recent efforts have been devoted to detect multiple sources in real-world situations, and the social influence of neighbors in the propagation is assumed to be identical. However, this assumption will result in inaccurate results as the infection state of a node is determined by its critical neighbors. In this paper, we fill this gap by capturing social influence of neighbors with structural properties in social networks. For instance, opinions are more likely to spread via closely connected friends within small groups. Here we propose a Motif-based Graph Convolutional Networks for Source Identification (MGCNSI) framework based on the GCN-based source identification approach. Specifically, different network motifs are used to capture different types of structural properties. Then each motif extracts the critical neighbors of a particular type, and a motif-based graph convolutional layer is constructed to aggregate critical neighbors for that motif. To adapt to underlying propagation mechanisms, an attention mechanism for aggregation is designed to automatically assign higher weights to more informative motifs. The empirical results demonstrate that MGCNSI outperforms several benchmark methods on both synthetic and real-world networks. The advantage is most obvious for networks with denser node neighborhoods, where MGCNSI can select critical neighbors from the larger neighbor sets. How the motifs can capture the social influence and the underlying critical paths of propagation is also illustrated.

show abstract

Section: Deep Learning On Graph Datamentioning

confidence: 99%

Identifying Multiple Propagation Sources With Motif-Based Graph Convolutional Networks for Social Networks

2023

View full text Add to dashboard Cite

show abstract

“…Recent years witness explosive growth in graph neural networks (GNNs) in pursuit of performance improvement of graph representation learning (Wu et al 2020;Liu et al 2022b;Lin et al 2022). GNNs are primarily designed for homogeneous graphs associated with a single type of nodes and edges, following a neighborhood aggregation scheme to capture structural information of a graph, where the representation of each node is computed by recursively aggregating the features of neighbor nodes (Kipf and Welling 2017).…”

Section: Introductionmentioning

confidence: 99%

Simple and Efficient Heterogeneous Graph Neural Network

Yang

Yan

Pan

et al. 2023

AAAI

View full text Add to dashboard Cite

Heterogeneous graph neural networks (HGNNs) have the powerful capability to embed rich structural and semantic information of a heterogeneous graph into node representations. Existing HGNNs inherit many mechanisms from graph neural networks (GNNs) designed for homogeneous graphs, especially the attention mechanism and the multi-layer structure. These mechanisms bring excessive complexity, but seldom work studies whether they are really effective on heterogeneous graphs. In this paper, we conduct an in-depth and detailed study of these mechanisms and propose the Simple and Efficient Heterogeneous Graph Neural Network (SeHGNN). To easily capture structural information, SeHGNN pre-computes the neighbor aggregation using a light-weight mean aggregator, which reduces complexity by removing overused neighbor attention and avoiding repeated neighbor aggregation in every training epoch. To better utilize semantic information, SeHGNN adopts the single-layer structure with long metapaths to extend the receptive field, as well as a transformer-based semantic fusion module to fuse features from different metapaths. As a result, SeHGNN exhibits the characteristics of a simple network structure, high prediction accuracy, and fast training speed. Extensive experiments on five real-world heterogeneous graphs demonstrate the superiority of SeHGNN over the state-of-the-arts on both accuracy and training speed.

show abstract