摘要
图回声状态网络（Graph Echo State Networks, GESN）已在图分类任务中展现出优异的性能与效率。然而，在半监督节点分类任务中，端到端训练的深度模型暴露出过平滑（over-smoothing）问题，导致模型对高同质性（high homophily）图结构产生偏差。本文首次在不同同质性程度的节点分类任务上评估了GESN的性能，并进一步分析了储备池半径（reservoir radius）的影响。实验结果表明，相较于需通过架构偏差进行特定调整的全训练深度模型，储备池模型在保持相当或更优准确率的同时，显著提升了计算效率。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| node-classification-on-actor | Graph ESN | Accuracy: 34.5 ± 0.8 |
| node-classification-on-chameleon | Graph ESN | Accuracy: 76.2±1.2 |
| node-classification-on-citeseer-full | Graph ESN | Accuracy: 74.5±2.1 |
| node-classification-on-cora-full-supervised | Graph ESN | Accuracy: 86.0±1.0 |
| node-classification-on-cornell | Graph ESN | Accuracy: 81.1±6.0 |
| node-classification-on-pubmed-full-supervised | Graph ESN | Accuracy: 89.2±0.3 |
| node-classification-on-squirrel | Graph ESN | Accuracy: 71.2±1.5 |
| node-classification-on-texas | Graph ESN | Accuracy: 84.3±4.4 |
| node-classification-on-wisconsin | Graph ESN | Accuracy: 83.3±3.8 |