A Gentle Introduction to Graph Neural Networks

Sánchez-Lengeling, Benjamín; Reif, Emily; Pearce, Adam; Wiltschko, Alexander B.

doi:10.23915/distill.00033

Cited by 155 publications

(198 citation statements)

References 8 publications

Supporting

Mentioning

173

Contrasting

Unclassified

Order By: Relevance

“…We perform experiments with all combinations of the aforementioned four datasets and three explainability techniques to quantify the explanation sharpening ability of SEEN. Since the ground-truth explanation (the binary labels whether each node is motif-participating) is available for the datasets due to their synthetic nature, we measure the area under the receiver operation characteristic curve (AUC-ROC) between the ground-truth and the obtained explanations, similar to Ying et al [12] and Sanchez-Lengeling et al [33]. Based on the measure, we calculate the difference in the explanation accuracy with or without SEEN on each combination.…”

Section: Discussionmentioning

confidence: 99%

“…To evaluate the explanation sharpening performance of SEEN, we adopt the following three GNN explainability techniques, which are compatible with explaining node classification tasks: sensitivity analysis (SA) [9,11,19], Grad*Input [27,33], and GradCAM [9,21,33]. Explainability methods based on gradients and features are particularly well-suited for SEEN due to their fast and notraining-required characteristics.…”

Section: Explainability Methodsmentioning

confidence: 99%

“…Grad*Input Similar to the SA on GNNs, Grad*Input [27,33] on GNNs calculates the explanation through element-wise multiplication of node features and their gradient over the prediction and the consequent summation over the feature dimension to generate node scores. Compared to the SA, the gradient from Grad*Input is directly multiplied to the input, and the absolute value is obtained after the multiplication.…”

Section: Sensitivity Analysismentioning

confidence: 99%

“…Node representations from intermediate message-passing layers are element-wisely multiplied with their gradients over prediction and summed into a single score value. A detailed formulation for applying GradCAM on GNNs varies from reported papers [9,33], such as the selection of the target layer, averaging the gradients node-wisely or taking the absolute value; we take an approach similar to the GradCAM(all) method applied by Sanchez-Lengeling et al [33], which multiplies the gradient and features element-wisely and averages the scores from each layer. Absolute values are obtained at the final stage of the explanation.…”

Section: Sensitivity Analysismentioning

confidence: 99%

See 3 more Smart Citations

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

Cho,

Oh,

Jeon

2021

Preprint

View full text Add to dashboard Cite

Explaining the foundations for predictions obtained from graph neural networks (GNNs) is critical for credible use of GNN models for real-world problems. Owing to the rapid growth of GNN applications, recent progress in explaining predictions from GNNs, such as sensitivity analysis, perturbation methods, and attribution methods, showed great opportunities and possibilities for explaining GNN predictions. In this study, we propose a method to improve the explanation quality of node classification tasks that can be applied in a post hoc manner through aggregation of auxiliary explanations from important neighboring nodes, named SEEN. Applying SEEN does not require modification of a graph and can be used with diverse explainability techniques due to its independent mechanism. Experiments on matching motif-participating nodes from a given graph show great improvement in explanation accuracy of up to 12.71% and demonstrate the correlation between the auxiliary explanations and the enhanced explanation accuracy through leveraging their contributions. SEEN provides a simple but effective method to enhance the explanation quality of GNN model outputs, and this method is applicable in combination with most explainability techniques.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Explainability Methodsmentioning

confidence: 99%

Section: Sensitivity Analysismentioning

confidence: 99%

Section: Sensitivity Analysismentioning

confidence: 99%

See 2 more Smart Citations

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

Cho,

Oh,

Jeon

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…We aim to explain the decision of an already trained GNN model (post-hoc interpretability) by attributing the reason for the underlying decision to either a subset of features or neighboring nodes, or both. Current approaches to explain GNNs [21,26,31] produce feature attributions in a post-hoc manner, but they suffer from some fundamental limitations.…”

mentioning

confidence: 99%

Zorro: Valid, Sparse, and Stable Explanations in Graph Neural Networks

Funke¹,

Khosla²,

Rathee³

et al. 2021

Preprint

View full text Add to dashboard Cite

With the ever-increasing popularity and applications of graph neural networks, several proposals have been made to interpret and understand the decisions of a GNN model. Explanations for a GNN model differ in principle from other input settings. It is important to attribute the decision to input features and other related instances connected by the graph structure. We find that the previous explanation generation approaches that maximize the mutual information between the label distribution produced by the GNN model and the explanation to be restrictive. Specifically, existing approaches do not enforce explanations to be predictive, sparse, or robust to input perturbations.In this paper, we lay down some of the fundamental principles that an explanation method for GNNs should follow and introduce a metric fidelity as a measure of the explanation's effectiveness. We propose a novel approach Zorro based on the principles from rate-distortion theory that uses a simple combinatorial procedure to optimize for fidelity. Extensive experiments on real and synthetic datasets reveal that Zorro produces sparser, stable, and more faithful explanations than existing GNN explanation approaches.

show abstract

Rational Design of Earth‐Abundant Catalysts toward Sustainability

Guo,

Haghshenas,

Jiao

et al. 2024

Advanced Materials

View full text Add to dashboard Cite

Catalysis is crucial for clean energy, green chemistry, and environmental remediation, but traditional methods rely on expensive and scarce precious metals. This review addresses this challenge by highlighting the promise of earth‐abundant catalysts and the recent advancements in their rational design. Innovative strategies such as physics‐inspired descriptors, high‐throughput computational techniques, and artificial intelligence (AI)‐assisted design with machine learning (ML) are explored, moving beyond time‐consuming trial‐and‐error approaches. Additionally, biomimicry, inspired by efficient enzymes in nature, offers valuable insights. This review systematically analyses these design strategies, providing a roadmap for developing high‐performance catalysts from abundant elements. Clean energy applications (water splitting, fuel cells, batteries) and green chemistry (ammonia synthesis, CO2 reduction) are targeted while delving into the fundamental principles, biomimetic approaches, and current challenges in this field. The way to a more sustainable future is paved by overcoming catalyst scarcity through rational design.

show abstract

A Gentle Introduction to Graph Neural Networks

Cited by 155 publications

References 8 publications

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

Zorro: Valid, Sparse, and Stable Explanations in Graph Neural Networks

Rational Design of Earth‐Abundant Catalysts toward Sustainability

Contact Info

Product

Resources

About