When Do GNNs Work: Understanding and Improving Neighborhood Aggregation

Xie, Yiqing; Sha, Li; Yang, Carl; Wong, Raymond Chi Wing

doi:10.24963/ijcai.2020/181

Cited by 52 publications

(30 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Convolutional neural networks (CNNs) applied on 2D/3D images defined on standard Euclidean grids (Deng et al, 2009;LeCun et al, 1989) are designed using 2D/3D rectangular convolutional kernels that slide across the images and map F in input feature maps to F out output feature maps. An extension of this application on data types in irregular domains such as graphs, is typically expressed using neighborhood aggregation (Corso et al, 2020;Xie et al, 2020) or message passing (Gilmer et al, 2017) schemes.…”

Section: Spiral Convolutionmentioning

confidence: 99%

Analyzing Brain Morphology in Alzheimer’s Disease Using Discriminative and Generative Spiral Networks

Azcona

Besson

et al. 2021

Preprint

View full text Add to dashboard Cite

Several patterns of atrophy have been identified and strongly related to Alzheimer's disease (AD) pathology and its progression. Morphological changes in brain shape have been identified up to ten years before clinical diagnoses of AD, making its early diagnosis more desirable. We propose novel geometric deep learning frameworks for the analysis of brain shape in the context of neurodegeneration caused by AD. Our deep neural networks learn low-dimensional shape descriptors of multiple neuroanatomical structures, instead of handcrafted features for each structure. A discriminative network using spiral convolution on 3D meshes is constructed for the in-vivo binary classification of AD from healthy controls (HCs) using a fast and efficient "spiral" convolution operator on 3D triangular mesh surfaces of human brain subcortical structures extracted from T1-weighted magnetic resonance imaging (MRI). Our network architecture consists of modular learning blocks using residual connections to improve overall classifier performance. In this work: (1) a discriminative network is used to analyze the efficacy of disease classification using input data from multiple brain structures and compared to using a single hemisphere or a single structure. It also outperforms prior work using spectral graph convolution on the same the same tasks, as well as alternative methods that operate on intermediate point cloud representations of 3D shapes. (2) Additionally, visual interpretations for regions on the surface of brain structures that are associated to true positive AD predictions are generated and fall in accordance with the current reports on the structural localization of pathological changes associated to AD. (3) A conditional generative network is also implemented to analyze the effects of phenotypic priors given to the model (i.e. AD diagnosis) in generating subcortical structures. The generated surface meshes by our model indicate learned morphological differences in the presence of AD that agrees with the current literature on patterns of atrophy associated to the disease. In particular, our inference results demonstrate an overall reduction in subcortical mesh volume and surface area in the presence of AD, especially in the hippocampus. The low-dimensional shape descriptors obtained by our generative model are also evaluated in our discriminative baseline comparisons versus our discriminative network and the alternative shape-based approaches.

show abstract

Section: Spiral Convolutionmentioning

confidence: 99%

Analyzing Brain Morphology in Alzheimer’s Disease Using Discriminative and Generative Spiral Networks

Azcona

Besson

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…A recent work [23] offers an important insight into the neighborhood aggregation of GNNs, i.e., the neighborhood aggregation is not always necessary and beneficial. Following this insight, we aim to deepen our understanding of the vulnerability of GNNs from an empirical study perspective.…”

Section: Related Workmentioning

confidence: 99%

“…There are several works in studying the expressive power of GNNs [10,17,23,25], including its depth, width, representational properties and limitations. Specifically, most of GNNs follow a neighborhood aggregation scheme, where the hidden Fig.…”

Section: Message Passing Graph Neural Networkmentioning

confidence: 99%

“…In addition, Xie et al [23] argue that the neighborhood aggregation operation is not always necessary and beneficial when the neighbors of a node are highly dissimilar. In this spirit, we propose two measurements of entropy to depict the chaos in node neighborhoods and use them to represent the predictability of a graph.…”

Section: Neighborhood Entropy Measurementsmentioning

confidence: 99%

“…Indeed, the following adversarial patterns are observed: High Neighborhood Entropy. Previously, Xie et al [23] mention that the neighborhood aggregation is not always necessary and beneficial for GNNs. Actually, it is consistent with our claim that the main reason for the vulnerability of GNNs is the non-robust message functions.…”

Section: Adversarial Patterns Of Attacks On Graph Datamentioning

confidence: 99%

See 2 more Smart Citations

Deep Insights into Graph Adversarial Learning: An Empirical Study Perspective

Peng

et al. 2021

Human Brain and Artificial Intelligence

View full text Add to dashboard Cite

Graph Neural Networks (GNNs) have shown to be vulnerable against adversarial examples in many works, which encourages researchers to drop substantial attention to its robustness and security. However, so far, the reasons for the success of adversarial attacks and the intrinsic vulnerability of GNNs still remain unclear. The work presented here outlines an empirical study to further investigate these observations and provide several insights. Experimental results, analyzed across a variety of benchmark GNNs on two datasets, indicate that GNNs are indeed sensitive to adversarial attacks due to its non-robust message functions. To exploit the adversarial patterns, we introduce two measurements to depict the randomness of node labels and features for a graph, noticing that the neighborhood entropy significantly increased under adversarial attacks. Furthermore, we find out that the adversarially manipulated graphs typically tend to be much denser and high-rank, where most of the dissimilar nodes are intentionally linked. And the stronger the attacks are, such as Metattack, the patterns are more apparent. To sum up, our findings shed light on understanding adversarial attacks on graph data and lead potential advancement in enhancing the robustness of GNNs.

show abstract

DMSPool: Dual Multi-Scale Pooling for Graph Representation Learning

Luo

et al. 2021

Database Systems for Advanced Applications

View full text Add to dashboard Cite

When Do GNNs Work: Understanding and Improving Neighborhood Aggregation

Cited by 52 publications

References 1 publication

Analyzing Brain Morphology in Alzheimer’s Disease Using Discriminative and Generative Spiral Networks

Analyzing Brain Morphology in Alzheimer’s Disease Using Discriminative and Generative Spiral Networks

Deep Insights into Graph Adversarial Learning: An Empirical Study Perspective

DMSPool: Dual Multi-Scale Pooling for Graph Representation Learning

Contact Info

Product

Resources

About