PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey on Graph Neural Networks

Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, Philip S. Yu

arXiv:1901.00596v4cs.LGstat.ML

TL;DR

Graph data are increasingly important but their irregular structure and complex relationships challenge conventional machine-learning methods, while existing GNN surveys cover only part of the field. This survey presents a comprehensive four-part taxonomy and synthesizes methods, applications, datasets, open-source codes, model assessment, and future directions.

  • Problem

    Existing GNN surveys examine limited subsets of methods, while graph data introduce irregular structures and complex relationships that challenge existing algorithms.

  • Method

    The survey organizes GNNs into recurrent, convolutional, autoencoder, and spatial-temporal categories, then reviews applications, datasets, codes, and model assessment.

  • Results

    The survey provides comprehensive coverage of GNN methods, applications, benchmark datasets, open-source codes, model assessment, and four future research directions.

  • Takeaways & Limitations

    The unified overview is intended to help researchers enter the field and enable experts to compare GNN models.

Abstract

from arXiv · show

Deep learning has revolutionized many machine learning tasks in recent years, ranging from image classification and video processing to speech recognition and natural language understanding. The data in these tasks are typically represented in the Euclidean space. However, there is an increasing number of applications where data are generated from non-Euclidean domains and are represented as graphs with complex relationships and interdependency between objects. The complexity of graph data has imposed significant challenges on existing machine learning algorithms. Recently, many studies on extending deep learning approaches for graph data have emerged. In this survey, we provide a comprehensive overview of graph neural networks (GNNs) in data mining and machine learning fields. We propose a new taxonomy to divide the state-of-the-art graph neural networks into four categories, namely recurrent graph neural networks, convolutional graph neural networks, graph autoencoders, and spatial-temporal graph neural networks. We further discuss the applications of graph neural networks across various domains and summarize the open source codes, benchmark data sets, and model evaluation of graph neural networks. Finally, we propose potential research directions in this rapidly growing field.

I. INTRODUCTION · II. BACKGROUND & DEFINITION · A. Background

The survey motivates graph neural networks (GNNs) as deep-learning extensions for increasingly prevalent non-Euclidean graph data and presents a comprehensive organization of their methods, applications, resources, and research directions. It also introduces foundational concepts, historical developments, and distinctions from network embedding and graph kernel methods.

  • I. INTRODUCTION: Graph data represent complex relationships and interdependencies in applications such as recommendation, drug discovery, and citation networks, challenging methods designed for Euclidean data.
  • I. INTRODUCTION: The survey proposes four GNN categories: recurrent, convolutional, graph autoencoder, and spatial-temporal graph neural networks.
  • I. INTRODUCTION: It provides a comprehensive review of representative models, comparisons, algorithms, benchmark data sets, open-source codes, and practical applications.
  • I. INTRODUCTION: The survey addresses limitations and proposes future directions concerning model depth, scalability trade-offs, heterogeneity, and dynamicity.
  • II. BACKGROUND & DEFINITION: Section II establishes the background by listing commonly used notations and defining graph-related concepts.
  • A. Background: Early GNN research progressed from neural networks on directed acyclic graphs to recurrent models that learn node representations through propagation.
  • A. Background: Graph convolutional methods form two streams—spectral-based and spatial-based—while network embedding preserves topology and node content for downstream graph analytics.

B. Definition … B. Frameworks

The paper defines graph structures, attributes, directedness, and spatial-temporal dynamics, then organizes GNNs into four categories and describes their output tasks, training paradigms, and representative architectures.

  • B. Definition: A graph G = (V, E) consists of nodes and directed edges, with adjacency matrix A and optional node and edge attributes.Node features are represented by X ∈ R^n×d, while edge features are represented by X_e ∈ R^m×c.
  • B. Definition: Undirected graphs are special directed graphs with inverse edge pairs and have symmetric adjacency matrices.
  • B. Definition: Spatial-temporal graphs are attributed graphs whose node attributes change dynamically over time, represented as G(t) = (V, E, X(t)).The time-varying node feature matrix satisfies X(t) ∈ R^n×d.
  • III. CATEGORIZATION AND FRAMEWORKS: The taxonomy divides GNNs into recurrent, convolutional, graph autoencoder, and spatial-temporal graph neural networks.The categories and representative architectures are summarized in Table II and illustrated in Figure 2.
  • A. Taxonomy of Graph Neural Networks (GNNs): RecGNNs learn node representations through recurrent information exchange until equilibrium, inspiring message passing in spatial-based ConvGNNs.ConvGNNs generalize convolution from grid data to graphs by aggregating a node’s features with those of its neighbors.
  • B. Frameworks: GNN outputs support node-level regression or classification, edge-level classification or link prediction, and graph-level classification.Node and edge tasks use learned node representations, whereas graph-level tasks typically require pooling and readout operations.
  • B. Frameworks: GNNs can use semi-supervised, supervised, or unsupervised end-to-end training depending on the task and available labels.Examples include semi-supervised node classification with graph convolutions and softmax, supervised graph classification with convolution, pooling, and readout, and unsupervised graph embedding.
  • B. Frameworks: Representative architectures stack graph convolution, pooling, readout, or autoencoder components to propagate messages, coarsen graphs, summarize representations, or reconstruct adjacency.Unsupervised alternatives include graph-convolutional encoders with decoders or negative sampling that distinguishes linked positive pairs from sampled negative pairs.

IV. RECURRENT GRAPH NEURAL NETWORKS

Recurrent graph neural networks reuse shared parameters to update node representations recurrently, originating in work focused on directed acyclic graphs. Later methods extend recurrence to general graphs, enforce or approximate convergence, and address scalability and training limitations.

  • RecGNN overview: RecGNNs apply the same parameters recurrently over graph nodes to extract high-level node representations, with early work mainly targeting directed acyclic graphs.Their shared recurrent layer updates node representations across the graph.
  • General-graph recurrence: GNN* extends recurrent models to acyclic, cyclic, directed, and undirected graphs by repeatedly exchanging neighborhood information until reaching a stable equilibrium.Its sum aggregation supports differing neighborhood sizes and unknown neighborhood orderings.
  • Convergence mechanisms: GNN* requires a contractive recurrent function for convergence, while GraphESN uses a randomly initialized, untrained contractive encoder before training its output layer on fixed node states.GraphESN updates node states until the global graph state converges.
  • Gated recurrence: GGNN uses a GRU recurrent function with a fixed number of steps, avoiding convergence constraints but learning parameters with BPTT and storing intermediate states for all nodes.These memory requirements can be problematic for large graphs.
  • Scalable recurrence: SSE improves scalability by asynchronously and stochastically updating sampled node batches, but it does not theoretically prove convergence to fixed points.Its stability mechanism averages historical and newly computed states.

V. CONVOLUTIONAL GRAPH NEURAL NETWORKS

Convolutional graph neural networks resolve cyclic node dependencies through a fixed stack of differently weighted layers rather than recurrent state iteration, making graph convolutions easier to combine with other neural networks.

  • Architectural distinction: ConvGNNs architecturally address cyclic mutual dependencies using a fixed number of layers with different weights in each layer.This contrasts with recurrent graph neural networks, which iterate node states under contractive constraints.
  • Advantages and adoption: Graph convolutions are more efficient and convenient to composite with other neural networks, contributing to the rapidly growing popularity of ConvGNNs.

A. Spectral-based ConvGNNs

Spectral-based ConvGNNs model graph signals through the normalized Laplacian’s eigenbasis and graph Fourier transform, with filter design distinguishing specific methods. Spectral CNN, ChebNet, CayleyNet, and GCN progressively address computational, localization, and frequency-selection considerations.

  • Spectral foundation: Spectral-based methods assume undirected graphs and use the normalized graph Laplacian, which is real symmetric positive semidefinite and admits an eigen-decomposition.The Laplacian is factored as L = UΛU^T, where U contains orthonormal eigenvectors and Λ contains eigenvalues.
  • Spectral foundation: The graph Fourier transform projects a node feature vector onto the Laplacian eigenvectors, while spectral convolution applies filtering in this transformed space.The transform is F(x) = U^T x and the inverse is F^-1(x̂) = Ux̂.
  • Filter designs: All spectral-based ConvGNNs share the same convolution definition, differing primarily in their choice of filter gθ.This filter choice determines how each method parameterizes or approximates spectral graph convolution.
  • Spectral CNN: Spectral CNN uses learnable diagonal spectral filters but faces graph-perturbation sensitivity, domain-dependent filters, and O(n^3) eigen-decomposition complexity.Its learned filters cannot be directly applied to graphs with different structures because the eigenbasis changes with graph perturbations.
  • ChebNet, CayleyNet, and GCN: ChebNet replaces exact spectral filtering with Chebyshev-polynomial approximations whose filters are localized in space and can extract local features independently of graph size.CayleyNet extends this approach with parametric rational complex functions for narrow frequency bands, while GCN introduces a first-order ChebNet approximation.

2 AD−1 2 x. (10) · 2 AD−1 2 )x. (11)

The section presents GCN’s parameter-constrained graph convolution, its multi-channel compositional extension, normalization for numerical stability, and neighborhood-aggregation interpretation. It also describes later variants that learn structural relations or combine adjacency and random-walk co-occurrence information.

  • 2 AD−1 2 )x. (11): GCN extends the convolution into a compositional layer to support multiple input and output channels.The modification is made to Equation 11.
  • 2 AD−1 2 )x. (11): The term 2 AD−1 2 empirically causes numerical instability, motivating GCN’s normalization trick.The trick replaces ¯A = In + D−1.
  • 2 AD−1 2 )x. (11): From a spatial perspective, GCN aggregates feature information from each node’s neighborhood, although it is introduced as a spectral-based method.Equation 12 is expressed using this neighborhood-aggregation interpretation.
  • 2 AD−1 2 )x. (11): GCN uses ˜A = A + In and ˜Dii = Pj ˜Aij in its normalized adjacency construction.These definitions specify the self-loop-augmented adjacency and its degree terms.
  • 2 AD−1 2 )x. (11): AGCN learns hidden structural relations through a residual adjacency matrix built from a learnable distance function on node features.This is presented as an incremental improvement over GCN using an alternative symmetric matrix.
  • 2 AD−1 2 )x. (11): DGCN uses two parameter-sharing layers with normalized adjacency and a PPMI matrix capturing node co-occurrence from sampled random walks.The PPMI matrix is defined using node and joint occurrence counts.
  • 2 AD−1 2 )x. (11): By ensembling dual graph-convolution outputs, DGCN encodes local and global structural information without stacking multiple graph-convolution layers.Its two layers share parameters and use complementary structural matrices.

B. Spatial-based ConvGNNs

Spatial-based ConvGNNs define graph convolutions from nodes’ spatial relations, typically propagating and aggregating neighborhood information through message passing. The section covers diffusion, shortest-path and partitioned neighborhoods, general message-passing frameworks, attention mechanisms, and scalable neighbor sampling.

  • Spatial-based graph convolution: Spatial-based ConvGNNs define graph convolutions from a node’s spatial relations, analogous to applying image filters over a pixel and its nearby neighbors.Images can be viewed as graphs whose pixels are nodes connected to nearby pixels.
  • Early spatial methods: NN4G learns graph dependencies compositionally, using independent layer parameters and summing neighborhood information while incrementally extending the neighborhood.CGMM maintains spatial locality while adding probabilistic interpretability, whereas NN4G uses an unnormalized adjacency matrix that may produce differently scaled hidden states.
  • Diffusion and structured neighborhoods: DCNN models graph convolution as diffusion through transition probabilities, concatenating representations from multiple propagation steps while distant-neighbor contributions diminish with transition-matrix powers.PGC-DGCNN increases distant-neighbor contributions using shortest-path adjacency matrices, but computing them can cost O(n^3) at maximum.
  • Message passing: MPNN provides a general spatial ConvGNN framework in which K-step message passing propagates information along edges, followed by node prediction or graph-level readout.GIN modifies the central-node weight with a learnable ϵ(k) to address limitations in distinguishing graph structures.
  • Attention and scalable training: GAT learns relative neighbor weights with attention rather than treating contributions as identical or predetermined, and multi-head attention improves expressive capability and node-classification performance over GraphSage.GraphSage samples a fixed number of neighbors and uses permutation-invariant aggregation, while GraphSage, Sto-GCN, and Cluster-GCN trade off time and memory differently during training.

C. Graph Pooling Modules

Graph pooling modules down-sample node features to produce smaller graph representations, reducing computational cost and helping avoid overfitting. Existing methods range from primitive mean/max/sum reduction and node rearrangement to learned hierarchical pooling, but effectiveness and computational complexity remain open challenges.

  • Pooling objectives: Pooling down-samples nodes after GNN feature generation to create smaller representations, reducing parameter size, computational demands, and overfitting risk.The strategy also supports permutation invariance.
  • Basic pooling: Mean/max/sum pooling is a fast and effective primitive for graph down-sampling.Its speed comes from computing the corresponding value within a pooling window.
  • Adaptive pooling: Set2Set addresses fixed-size embedding limitations by using an LSTM to integrate order-dependent information into memory before reduction.Its memory increases with input graph size, avoiding information loss from direct reduction.
  • Node-rearrangement pooling: ChebNet and DGCNN rearrange nodes before pooling, using balanced binary trees or structural roles to impose meaningful order.DGCNN’s SortPooling additionally unifies graph size by truncating or extending the ordered node features.
  • Learned pooling: DiffPool learns cluster assignments from node features and topology to generate hierarchical graph representations, but its dense pooled graphs incur O(n^2) computational complexity.SAGPool likewise considers node features and topology through self-attention.
  • Open challenges: Improving pooling effectiveness and computational complexity remains an open research question.Pooling is described as essential for reducing graph size.

D. Discussion of Theoretical Aspects

This section examines theoretical foundations of graph neural networks, including receptive fields, complexity, graph isomorphism, permutation behavior, and universal approximation. It summarizes how these properties constrain or characterize GNN representations and functions.

  • Shape of receptive field: Spatial graph convolutional layers expand a node’s receptive field one hop per layer, and finitely many layers can cover every node in the graph.Micheli proves that a finite number of spatial graph convolutional layers suffices for each node’s receptive field to encompass the entire graph.
  • VC dimension: The VC dimension of a GNN* is O(p4n2) with sigmoid or tangent hyperbolic activation and O(p2n) with piecewise polynomial activation.Here, p is the number of model parameters and n is the number of nodes.
  • Graph isomorphism: GNN embeddings can support Weisfeiler-Lehman identification of non-isomorphic graphs, but common GCN and GraphSage models cannot distinguish all different graph structures.Xu et al. further analyze how aggregation functions and related components affect this limitation.
  • Equivariance and invariance: Node-level GNNs must be permutation equivariant, satisfying f(QAQT , QX) = Qf(A, X), whereas graph-level GNNs must be permutation invariant, satisfying f(QAQT , QX) = f(A, X).Achieving these properties requires GNN components to be invariant to node orderings.
  • Universal approximation: A RecGNN can approximate any function that preserves unfolding equivalence to any degree of precision.The section notes that universal approximation of GNNs has been studied less extensively than that of one-hidden-layer feedforward networks.

VI. GRAPH AUTOENCODERS · A. Network Embedding

Graph autoencoders map nodes into latent representations and decode graph information, supporting network embedding and graph generation. For network embedding, surveyed methods preserve structural or feature information through reconstruction, variational regularization, contrastive objectives, or sequence-based modeling.

  • VI. GRAPH AUTOENCODERS: Graph autoencoders map nodes into a latent feature space and decode graph information for network embedding or graph generation.The survey divides its review of GAEs into network embedding and graph generation.
  • A. Network Embedding: Network embeddings are low-dimensional node vectors that preserve topology, learned by encoding nodes and decoding structural information such as PPMI or adjacency matrices.The encoder extracts embeddings, while the decoder enforces preservation of graph topology.
  • A. Network Embedding: DNGR reconstructs the PPMI matrix with a stacked denoising autoencoder, while SDNE jointly preserves first-order and second-order node proximity.Both earlier approaches use multi-layer perceptrons to construct graph autoencoders.
  • A. Network Embedding: GAE* extends structural-only embedding by using graph convolutional layers to encode node structure and features, then reconstructing the adjacency matrix.Its training minimizes negative cross entropy between the real and reconstructed adjacency matrices.
  • A. Network Embedding: VGAE addresses potential overfitting from adjacency reconstruction by learning a data distribution through variational lower-bound optimization.Its objective encourages the empirical distribution q(Z|X, A) to approach a Gaussian prior p(Z).
  • A. Network Embedding: GraphSage preserves relational information with negative sampling, encouraging neighboring nodes to have similar representations and distant nodes to have dissimilar representations.The loss samples distant nodes from a negative sampling distribution.
  • A. Network Embedding: DGI maximizes local mutual information to make local embeddings capture global structure and shows a distinct experimental improvement over GraphSage.These methods generally learn embeddings by solving link prediction, whose graph sparsity creates far fewer positive than negative node pairs.
  • A. Network Embedding: Sequence-based approaches alleviate graph sparsity by converting graphs into random-permutation or random-walk sequences for deep learning models such as LSTMs.DRNE uses an LSTM over degree-ordered neighbor sequences while avoiding permutation sensitivity; NetRA uses rooted random walks and adversarial regularization.

B. Graph Generation · VII. SPATIAL-TEMPORAL GRAPH NEURAL NETWORKS

Graph autoencoders generate graphs either sequentially or globally, trading structural fidelity and scalability in different ways. Spatial-temporal graph neural networks model dynamic node inputs and spatial dependencies using recurrent, convolutional, or adaptive-graph approaches.

  • B. Graph Generation: Graph autoencoders learn graph-generative distributions by encoding graphs into hidden representations and decoding graph structures, especially for molecular graph generation.They generate graphs sequentially or globally, with applications in drug discovery.
  • B. Graph Generation: Sequential methods generate nodes and edges step by step, including SMILES-based models that use deep CNNs and RNNs for encoding and decoding.General graph methods iteratively add nodes and edges until a criterion is satisfied.
  • B. Graph Generation: Global methods output graphs simultaneously; GraphVAE models node and edge existence as independent random variables and decodes adjacency, node attributes, and edge attributes.Controlling connectivity, validity, and node compatibility remains challenging.
  • B. Graph Generation: Sequential generation can lose structural information from cycles, whereas global generation is not scalable to large graphs because the GAE output space reaches O(n2).DeepGMG makes sequential decisions conditioned on node states and the growing graph state updated by a RecGNN.
  • VII. SPATIAL-TEMPORAL GRAPH NEURAL NETWORKS: STGNNs capture dynamic graph behavior by modeling changing node inputs while assuming interdependency between connected nodes.Traffic networks exemplify this setting, with sensor edges weighted by distances between sensors.
  • VII. SPATIAL-TEMPORAL GRAPH NEURAL NETWORKS: RNN-based STGNNs apply graph convolutions to inputs and hidden states, with GCRN combining LSTM and ChebNet and DCRNN combining diffusion convolution with GRU.DCRNN uses an encoder-decoder framework to predict future K steps of node values.
  • VII. SPATIAL-TEMPORAL GRAPH NEURAL NETWORKS: Structural-RNN separates temporal processing into node-RNNs and edge-RNNs, while node-RNN inputs incorporate edge-RNN outputs to capture spatial information.It predicts node labels at each time step.
  • VII. SPATIAL-TEMPORAL GRAPH NEURAL NETWORKS: CNN-based approaches avoid recurrent propagation through interleaved 1D-CNN and graph-convolution layers, offering parallel computing, stable gradients, and low memory requirements.Graph WaveNet additionally learns a self-adaptive adjacency matrix from node embeddings and can perform well without a supplied adjacency matrix, while GaAN learns dynamic spatial dependencies with attention.

VIII. APPLICATIONS … C. Practical Applications

The survey organizes GNN resources and applications by summarizing benchmark data sets, evaluation practices, implementations, and practical uses across graph-related tasks and domains. It highlights broad applicability while noting important limitations in evaluation comparability.

  • A. Data Sets: The survey summarizes selected benchmark data sets in four groups: citation networks, biochemical graphs, social networks, and others.More details are provided in Supplementary Material A.
  • B. Evaluation & Open-source Implementations: Node and graph classification are common evaluation tasks, but reported results and experimental settings are not necessarily rigorous or unified.Node classification commonly uses train/validation/test splits and reports average accuracy or F1 over multiple runs, while graph classification often uses 10-fold cross-validation.
  • B. Evaluation & Open-source Implementations: Open-source implementations support baseline experiments, including PyTorch Geometric and the Deep Graph Library for implementing many GNNs.The survey provides hyperlinks to reviewed model implementations in Supplementary Material C.
  • C. Practical Applications: GNNs address general graph tasks including node and graph classification, network embedding, graph generation, spatial-temporal forecasting, node clustering, link prediction, and graph partitioning.These capabilities motivate applications across multiple research domains.
  • C. Practical Applications: In computer vision, GNNs support scene graph generation, point-cloud classification, and action recognition by modeling semantic, topological, or skeletal relationships.Additional directions include human-object interaction, few-shot image classification, semantic segmentation, visual reasoning, and question answering.
  • C. Practical Applications: In natural language processing, GNNs use document, word, or syntactic relationships for text classification and graph-to-sequence or sequence-to-graph learning.Applications include generating sentences from semantic graphs and constructing semantic or knowledge graphs from sentences.
  • C. Practical Applications: GNNs model traffic networks, recommender systems, molecular graphs, and diverse other problems such as program reasoning, brain networks, event detection, and combinatorial optimization.Traffic forecasting uses spatial-temporal graphs; recommendation is cast as link prediction; molecular graphs represent atoms as nodes and bonds as edges.

IX. FUTURE DIRECTIONS · X. CONCLUSION

The survey identifies four future directions for GNNs arising from graph complexity: model depth, scalability trade-offs, heterogeneity, and dynamicity. It concludes by organizing GNN research into four categories and reviewing methods and applications.

  • IX. FUTURE DIRECTIONS: The authors identify four future directions for GNNs because graph complexity continues to create challenges despite their demonstrated power.The four directions are model depth, scalability trade-offs, heterogeneity, and dynamicity.
  • IX. FUTURE DIRECTIONS: Model depth remains problematic because ConvGNN performance can drop dramatically as graph convolutional layers increase, while repeated convolutions push adjacent-node representations closer together.With infinitely many graph convolutional layers, all node representations theoretically converge to a single point.
  • IX. FUTURE DIRECTIONS: Scalability methods trade graph completeness for efficiency: sampling can omit influential neighbors, whereas clustering can remove distinct structural patterns.The survey frames balancing algorithm scalability with graph integrity as an open research direction.
  • IX. FUTURE DIRECTIONS: Most current GNNs assume homogeneous graphs, making them difficult to apply directly to graphs with varied node or edge types and multimodal inputs.The survey calls for methods that handle heterogeneous graphs containing inputs such as images and text.
  • IX. FUTURE DIRECTIONS: Dynamic graphs require new graph convolutions because nodes, edges, and their inputs may change over time, while few STGNNs address changing spatial relations.Spatial-temporal graph neural networks partly address graph dynamicity but do not broadly handle dynamic spatial relations.
  • X. CONCLUSION: The survey groups GNNs into recurrent, convolutional, autoencoder, and spatial-temporal categories, then reviews and compares methods within and across these categories.It also introduces a wide range of GNN applications and summarizes datasets and open-source codes.

APPENDIX … C. Open-source Implementations

The appendix describes benchmark datasets spanning citation, biochemical, biological, social, image, and spatial-temporal graphs, summarizes reported node-classification results, and provides source-code hyperlinks for reviewed GNN models.

  • A. Data Set: Citation networks represent papers, authors, citations, authorship, and co-authorship, and commonly support node classification, link prediction, and node clustering evaluation.Although directed, citation networks are often treated as undirected for model evaluation; Cora, Citeseer, and Pubmed are popular datasets.
  • A. Data Set: Biochemical graph benchmarks represent atoms as nodes and chemical bonds as edges for graph classification tasks.NCI-1 and NCI-9 contain 4110 and 4127 compounds, while MUTAG contains 188 nitro compounds with categorical labels.
  • A. Data Set: The PPI benchmark contains 24 biological graphs whose nodes are proteins, edges are interactions, and node labels encode biological states.Each graph corresponds to one human tissue.
  • A. Data Set: Social-network benchmarks include BlogCatalog, which links bloggers by social relationships, and Reddit, which links posts commented on by the same user.BlogCatalog labels represent personal interests, whereas Reddit labels indicate discussion communities.
  • A. Data Set: Other benchmarks include graph-constructed MNIST images and METR-LA, a spatial-temporal traffic graph collected from 207 Los Angeles County highway sensors.MNIST uses 8-nearest-neighbors based on pixel locations; METR-LA contains four months of traffic data.
  • B. Reported Experimental Results for Node Classification: Table VII summarizes experimental results for node classification under a standard train/valid/test split, using accuracy for Cora, Citeseer, and Pubmed and micro-averaged F1 for PPI and Reddit.The split summary is stated separately, while the table caption identifies the metrics used for each dataset group.
  • C. Open-source Implementations: The survey summarizes open-source implementations of the reviewed graph neural networks and provides hyperlinks to their source code in Table VIII.This section serves as an implementation reference for the GNN models covered by the survey.
Loading 1901.00596v4…