返回主页

Marked in the browser
# CGIPool: Graph Pooling via Coarsened Graph Infomax [论文下载](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190908942.pdf) >SIGIR 2021 ## 论文解读 ### 论文核心思想 通过最大化原图与池化图的互信息,使得原图的特征与池化后图的特征更加接近。因此既能达成池化的效果,又能保证池化前后特征不变。 ### 总体框架 ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190909047.png) ### Positive and Negative Coarsening Module (1)计算real和fake粗化图分数: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190911876.png) (2) real和fake粗化图都选择前k个分数最大的节点: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190912166.png) (3)得到邻接矩阵和特征矩阵: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190912838.png) (4)原图、real、fake粗化图经过共享编码器: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190913232.png) (5) 互信息最大化: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190913265.png) (6) 融合real,fake粗化图: ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190914218.png) ### 实验结果 (1)图分类实验 ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190915359.png) (2)消融实验 ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190915335.png) ![](https://hexo-img.obs.cn-east-3.myhuaweicloud.com/img/202306190915744.png)