A Hyperbolic-to-Hyperbolic Graph Convolutional Network

Dai, Jindou; Wu, Yuwei; Gao, Zhi; Jia, Yingmin

doi:10.1109/cvpr46437.2021.00022

Cited by 39 publications

(34 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…NNs methods are also baseline model and they only exploit the feature of nodes, neglecting the graph-structured data. For GCNs and HYP GCNs, the stateof-art models, like H2H-GCN [12] and GCN [5], GRAPHSAGE [4], graph attention networks (GAT) [39], HGCN [8] and simplified graph convolution (SGC) [14].…”

Section: Node Classificationmentioning

confidence: 99%

Adaptive Frequency-Based Fully Hyperbolic Graph Neural Networks

Wei

Ping

Mei

2022

Preprint

View full text Add to dashboard Cite

Graph Convolutional Networks (GCNs) have attracted broad attention from industry and academia, for which GCNs have demonstrated powerful ability to model the irregular data, e.g., skeletal data and graph-structured data. The most existing effective model may be the fully hyperbolic graph neural network. However, it involves a large number of parameters, thus consuming considerable computing resources. In this paper, we propose a model based on adaptive frequency filter and corresponding optimizer in hyperbolic space. The adaptive frequency can learn the different frequency components of the embeddings of the nodes in graph, which adaptively adjust the beneficial signals of high-frequency and low-frequency. And the optimizer is based on a subset of the orthogonal constraint, which is dedicated for the adaptive frequency with less parameters. Consequently, our model need only to optimize the less parameters in hyperbolic space and meanwhile prevent the distortion caused by conventional manifold GCN. Experimental results show that our method achieves substantial improvements and outperforms the state-of-the-art performance in terms of node classification.

show abstract

Section: Node Classificationmentioning

confidence: 99%

Adaptive Frequency-Based Fully Hyperbolic Graph Neural Networks

Wei

Ping

Mei

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Another line of methods focuses on global graph representation learning, which can be directly used for graph classification tasks [52,55]. The key notion is developing a multi-scale feature extraction pipeline to leverage different levels of topological information in graphs [7,19,46,48]. GIN is proposed as a graph classification method that is as powerful as the Weisfeiler-Lehman Isomorphism test [44].…”

Section: Related Work 21 Graph Neural Networkmentioning

confidence: 99%

“…The coefficient 𝜆 is selected from [0.005, 0.01, 0.02, 0.03]. For each dataset, the upper limit of BFS 𝛽 is selected from [3,5,7]. For larger datasets, the number of augmented edges is selected from [3,5,7], and from [1,2,3] for other datasets.…”

Section: Implementation Detailsmentioning

confidence: 99%

Supervised Contrastive Learning with Structure Inference for Graph Classification

Jia¹,

Ji²,

Lei³

2022

Preprint

View full text Add to dashboard Cite

Advanced graph neural networks have shown great potentials in graph classification tasks recently. Different from node classification where node embeddings aggregated from local neighbors can be directly used to learn node labels, graph classification requires a hierarchical accumulation of different levels of topological information to generate discriminative graph embeddings. Still, how to fully explore graph structures and formulate an effective graph classification pipeline remains rudimentary. In this paper, we propose a novel graph neural network based on supervised contrastive learning with structure inference for graph classification. First, we propose a data-driven graph augmentation strategy that can discover additional connections to enhance the existing edge set. Concretely, we resort to a structure inference stage based on diffusion cascades to recover possible connections with high node similarities. Second, to improve the contrastive power of graph neural networks, we propose to use a supervised contrastive loss for graph classification. With the integration of label information, the one-vs-many contrastive learning can be extended to a many-vs-many setting, so that the graph-level embeddings with higher topological similarities will be pulled closer. The supervised contrastive loss and structure inference can be naturally incorporated within the hierarchical graph neural networks where the topological patterns can be fully explored to produce discriminative graph embeddings. Experiment results show the effectiveness of the proposed method compared with recent state-of-the-art methods.

show abstract

“…The hyperbolic space has gained traction in deep learning literature for representing tree-like structures and taxonomies [18,20,29,30,36,38,47], text [2,42,55], and graphs [4,8,12,22,26,50]. Hyperbolic alternatives have been proposed for various network layers, from intermediate layers [17,39] to classification layers [3,11,17,39].…”

Section: Hyperbolic Deep Learningmentioning

confidence: 99%

“…Foundational work showed that hyperbolic manifolds are able to embed hierarchies and tree-like structures with minimal distortion [29]. Follow up work has demonstrated the benefits of hyperboles for various tasks with latent hierarchical structures, from text embedding [42,55] to graph inference [8,12,22]. Notably, Khrulkov et al [19] showed that hyperbolic embeddings also have profound connections to visual data, due to latent hierarchical structures present in vision datasets.…”

Section: Introductionmentioning

confidence: 99%

Hyperbolic Image Segmentation

GhadimiAtigh¹,

Schoep²,

Acar³

et al. 2022

Preprint

View full text Add to dashboard Cite

Figure 1. Two examples of insights that come for free with Hyperbolic Image Segmentation. For both examples, each black dot denotes a pixel embedding in hyperbolic space. Left (Pascal VOC): next to per-pixel classification, the distance to the origin in hyperbolic space provides a free measure of uncertainty. Right (COCO-Stuff-10k): the hyperbolic positioning of pixels even allows us to pinpoint interiors and edges of objects, as indicated by the colored boxes and their corresponding pixels in the segmentation map. Other benefits of hyperbolic embeddings for segmentation include zero-label generalization and better performance in low-dimensional embedding spaces.

show abstract

A Hyperbolic-to-Hyperbolic Graph Convolutional Network

Cited by 39 publications

References 19 publications

Adaptive Frequency-Based Fully Hyperbolic Graph Neural Networks

Adaptive Frequency-Based Fully Hyperbolic Graph Neural Networks

Supervised Contrastive Learning with Structure Inference for Graph Classification

Hyperbolic Image Segmentation

Contact Info

Product

Resources

About