摘要
在这项工作中，我们致力于将卷积神经网络（CNNs）从低维规则网格（图像、视频和语音在此类网格上表示）推广到高维不规则域，例如社交网络、脑连接组或词嵌入（这些域由图表示）。我们提出了在谱图理论框架下对CNNs进行公式化的方案，该理论提供了必要的数学背景和高效的数值方法来设计图上的快速局部卷积滤波器。重要的是，所提出的技术具有与经典CNN相同的线性计算复杂度和常数学习复杂度，同时适用于任何图结构。通过在MNIST和20NEWS数据集上的实验，证明了这一新型深度学习系统能够学习图上的局部、平稳和组合特征。
代码仓库
mdeff/cnn_graph
官方
tf
GitHub 中提及
hazdzz/ChebyNet
pytorch
GitHub 中提及
mdeff/paper-cnn-graph-nips2016
GitHub 中提及
ajbisberg/gcn
tf
GitHub 中提及
selmiss/gp-tlstgcn
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-property-prediction-on-ogbg-molpcba | ChebNet | Ext. data: No Number of params: 1475003 Test AP: 0.2306 ± 0.0016 Validation AP: 0.2372 ± 0.0018 |
| node-classification-on-citeseer | ChebNet | Accuracy: 69.8% |
| node-classification-on-citeseer-05 | ChebyNet | Accuracy: 45.3% |
| node-classification-on-citeseer-1 | ChebyNet | Accuracy: 59.4% |
| node-classification-on-citeseer-with-public | ChebyNet | Accuracy: 70.1% |
| node-classification-on-cora | ChebNet | Accuracy: 81.2% |
| node-classification-on-cora-05 | ChebyNet | Accuracy: 33.9% |
| node-classification-on-cora-1 | ChebyNet | Accuracy: 44.2% |
| node-classification-on-cora-3 | ChebyNet | Accuracy: 62.1% |
| node-classification-on-cora-with-public-split | ChebyNet | Accuracy: 78.0% |
| node-classification-on-pubmed | ChebNet | Accuracy: 74.4% |
| node-classification-on-pubmed-003 | ChebyNet | Accuracy: 45.3% |
| node-classification-on-pubmed-005 | ChebyNet | Accuracy: 48.2% |
| node-classification-on-pubmed-01 | ChebyNet | Accuracy: 55.2% |
| node-classification-on-pubmed-with-public | ChebyNet | Accuracy: 69.8% |
| skeleton-based-action-recognition-on-sbu | ChebyNet | Accuracy: 96.00% |