Scaling Up Graph Neural Networks Via Graph Coarsening

Huang, Zengfeng; Zhang, Shengzhong; Xi, Chong; Zhou, Min

doi:10.1145/3447548.3467256

Cited by 49 publications

(17 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, another possible solution to mitigate the cost of eigendecomposition of S is to use graph coarsening to reduce the size of graph or graph partition to partition a large graph into several small graphs without losing much information. Several works have proposed on this topics [3,12]. Worth note that graph partition is also used in a well-known work Cluster-GCN [5] which focuses on scaling up GCNs to a large graph with millions nodes.…”

Section: More Discussion About Limitationsmentioning

confidence: 99%

See 1 more Smart Citation

EIGNN: Efficient Infinite-Depth Graph Neural Networks

Liu¹,

Kawaguchi²,

Hooi³

et al. 2022

Preprint

View full text Add to dashboard Cite

Graph neural networks (GNNs) are widely used for modelling graph-structured data in numerous applications. However, with their inherently finite aggregation layers, existing GNN models may not be able to effectively capture long-range dependencies in the underlying graphs. Motivated by this limitation, we propose a GNN model with infinite depth, which we call Efficient Infinite-Depth Graph Neural Networks (EIGNN), to efficiently capture very long-range dependencies. We theoretically derive a closed-form solution of EIGNN which makes training an infinite-depth GNN model tractable. We then further show that we can achieve more efficient computation for training EIGNN by using eigendecomposition. The empirical results of comprehensive experiments on synthetic and real-world datasets show that EIGNN has a better ability to capture long-range dependencies than recent baselines, and consistently achieves state-of-the-art performance. Furthermore, we show that our model is also more robust against both noise and adversarial perturbations on node features.35th Conference on Neural Information Processing Systems (NeurIPS 2021), virtual.

show abstract

Section: More Discussion About Limitationsmentioning

confidence: 99%

“…The time complexity of a plain full eigendecomposition of S is O(n 3 ). In some cases where the complexity of the preprocessing is of importance, we can consider using truncated eigendecomposition or graph coarsening [12] to mitigate the cost. Regarding this, we provide more discussions in Appendix D.…”

Section: Comparison With Other Gnns With Infinite Depthmentioning

confidence: 99%

EIGNN: Efficient Infinite-Depth Graph Neural Networks

Liu¹,

Kawaguchi²,

Hooi³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…The dataset sources are also supplemented for reproducibility. The research benchmark datasets used for evaluation in 3 The default values of 𝛼 and 𝛽 are 1 this work are Disease [4], Twitch_en [22], CoauthorCS [23] and two benchmark citation networks, Citeseer [37] and Cora [9]. In Disease, the label of a node indicates whether the node is infected by a disease or not, and the node features associate the susceptibility to the disease.…”

Section: Experiments 31 Experimental Setupmentioning

confidence: 99%

BSAL: A Framework of Bi-component Structure and Attribute Learning for Link Prediction

Li,

Zhou,

Zhang

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Given the ubiquitous existence of graph-structured data, learning the representations of nodes for the downstream tasks ranging from node classification, link prediction to graph classification is of crucial importance. Regarding missing link inference of diverse networks, we revisit the link prediction techniques and identify the importance of both the structural and attribute information. However, the available techniques either heavily count on the network topology which is spurious in practice, or cannot integrate graph topology and features properly. To bridge the gap, we propose a bicomponent structural and attribute learning framework (BSAL) that is designed to adaptively leverage information from topology and feature spaces. Specifically, BSAL constructs a semantic topology via the node attributes and then gets the embeddings regarding the semantic view, which provides a flexible and easy-to-implement solution to adaptively incorporate the information carried by the node attributes. Then the semantic embedding together with topology embedding are fused together using attention mechanism for the final prediction. Extensive experiments show the superior performance of our proposal and it significantly outperforms baselines on diverse research benchmarks.

show abstract

“…GNNbased methods formulate the link prediction as the binary classification problem in which the explicit node feature could be incorporated [9,24,26]. These techniques are typically evaluated on a limited number of regular benchmark datasets such as collaboration network [19] or citation networks [8]. However, many real-world scenarios, such as protein interaction networks [20], knowledge graphs [16], disease spreading network [4], transport networks [1], and social networks [3], are usually sparse or hierarchical in nature, which calls for special attentions [3,15,22].…”

Section: Base Stationmentioning

confidence: 99%

TeleGraph: A Benchmark Dataset for Hierarchical Link Prediction

Zhou¹,

Li²,

Yang³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Link prediction is a key problem for network-structured data, attracting considerable research efforts owing to its diverse applications. The current link prediction methods focus on general networks and are overly dependent on either the closed triangular structure of networks or node attributes. Their performance on sparse or highly hierarchical networks has not been well studied. On the other hand, the available tree-like benchmark datasets are either simulated, with limited node information, or small in scale. To bridge this gap, we present a new benchmark dataset TeleGraph, a highly sparse and hierarchical telecommunication network associated with rich node attributes, for assessing and fostering the link inference techniques. Our empirical results suggest that most of the algorithms fail to produce satisfactory performance on a nearly tree-like dataset, which calls for special attention when designing or deploying the link prediction algorithm in practice.

show abstract

Scaling Up Graph Neural Networks Via Graph Coarsening

Cited by 49 publications

References 16 publications

EIGNN: Efficient Infinite-Depth Graph Neural Networks

EIGNN: Efficient Infinite-Depth Graph Neural Networks

BSAL: A Framework of Bi-component Structure and Attribute Learning for Link Prediction

TeleGraph: A Benchmark Dataset for Hierarchical Link Prediction

Contact Info

Product

Resources

About