Table 1 GNN papers with their performance
Refs. | Application area | Dataset used | Model applied | Summary | Performance evaluation |
|---|---|---|---|---|---|
[13] 2018 | Link prediction | Usair NS PB Yeast C.ele Power Router And E.coli | GNN | They defend the use of prediction heuristics to learn from local enclosing subgraphs | Area Under Curve: Usair—97.09 ± 0.70 NS—97.71 ± 0.93 PB—95.01 ± 0.34 Yeast—97.20 ± 0.64 C.ele—89.54 ± 2.04 Power—84.18 ± 1.82 Router—95.68 ± 1.22 E.coli—97.22 ± 0.28 |
[14] 2018 | Solving matrix equations | Own examples | Hybrid GNN GNN + ZNN (Zhang Neural Network) | They solve the matrix equations BX = D and XC = D in time-invariant cases | The global convergence rate has improved by taking their example |
[15] 2018 | Solving matrix equations | Explains theorems with their examples | Gradient-based neural dynamics (GND) | Solve matrix equation AXB = D | The global convergence rate has improved by taking their example |
[16] 2019 | Link forecast Recommendation Node Clustering and Node Classification | Academic I (A-I) Academic II (A-II) Movies Review (R-I) Cds Review | Hetgnn | Hetgnn considered combining heterogeneous types, type-based neighbors, and heterogeneous node contents | AUC: Multi-label classification—0.978 Node clustering – 0.901 |
[17] 2019 | Social Recommendation | Ciao and Epinions | Graphrec | They predict ratings and provide interactions and opinions on the user-item graph | RMSE: Ciao—0.9794 Epinions—0.8168 |
[18] 2019 | Chinese-named entity recognition | Ontonotes MSRA Resume | Lexicon-based GNN | Chinese NER is achieved as a graph node classification using a vocabulary to build a graph neural network | 74.89 93.46 60.21 95.37 |
[19] 2019 | Taxable detection & structure recognition | UW3, UNLV, ICDAR 2013 | CNN + GNN | The best networks for detecting representative visual features are convolutional neural networks, whereas the best networks for quick message transfer between vertices are graph networks. With the help of the gather operation, we have demonstrated how to integrate these two skills | 68.5 |
[20] 2019 | Link prediction Pair-wise node classification | Grid Communities PPI | P-GNN Point of view GNN | To compute node embeddings that contain node positional information while maintaining inductive capability and leveraging node attributes, they introduce a new class of GNN | AUC: 0.940 ± 0.027 0.985 ± 0.008 0.808 ± 0.003 |
[21] 2020 | Time-series Forecasting | METR-LA PEMS-BAY PEMS07 PEMS03 PEMS04 PEMS08 Solar Electricity ECG5000 COVID-19 | Spectral Temporal GNN Stemgnn) | Â | RMSE: 5.06 2.48 4.01 21.64 32.15 24.93 0.07 0.06 0.07 19.3 |
[22] [2020] | Citation Network | Cora, Citeseer, Pubmed, and NELL | Continuous GNN(CGNN) | Enable continuous instances to be handled by existing discrete graph neural networks by describing the evolution of node representations with ODE | 82.1 ± 1.3 72.9 ± 0.9 82.7 ± 1.4 73.1 ± 0.9 |
[23] 2020 | Node representation visualization | Cora, Citeseer, Pubmed Coauthors | Differentiable group normalization (DGN), simple graph convolution networks (SGC) | They propose group distance ratio and instance information gain as two over-smoothing metrics based on graph architectures | 80.2% 58.2% 76.2% 85.8% |
[24] 2020 | Medical | MUTAG PTC COX2 PROTEINS NCI1 | Implicit graph neural network (IGCN) | They outline a Perron-Frobenius hypothesis necessary condition for very well and a projected gradient descent training approach | 89.3 ± 6.7 70.1 ± 5.6 86.9 ± 4.0 77.7 ± 3.4 80.5 ± 1.9 |
[25] 2021 | Text classification | IMDB webkb R52 R8 AG_news | Deep Attention Diffusion Graph Neural Network (DADGNN) | Proposes an attention diffusion technique that captures non-direct-neighbor context information in a single layer and decouples the required GNN training processes (representation transformation and propagation) | 88.49 ± 0.59 90.92 ± 0.42 95.16 ± 0.22 98.15 ± 0.16 92.24 ± 0.36 |
[26] 2021 | Medicines | COLL, MD17, and OC20 | Neural Network For Geometric Messages Passing | Gemnet uses effective bilinear layers and symmetric message passing | 34%, 41%, and 20% |
[27] 2022 | Feature extraction | Pavia University Salinas Houston 2013 | Deep Hybrid Multi-Graph Neural Network (DHMG) | To reduce the noise in the graph, they created a unique ARMA filter and implemented it recursively | 97.81 ± 0.82 98.33 ± 0.28 93.31 ± 0.65 |
[28] (2023) | Traffic Prediction | AIS data and global port geospatial data | GAT | Research on Multi-Port Ship Traffic Prediction Method Based on Spatiotemporal Graph Neural Networks | Around 90% |
[29] (2023) | Traffic Forecasting | PEMS03, PEMS04, PEMS07, PEMS08 | GNN | Hybrid GCN and branch-and-bound optimization for traffic flow forecasting | 0.58 0.63 0.63 0.73 |