GNN 应用的案例

sunr 发表于 2022-11-05 16:02:01
求解graph embedding到底怎么理解，感觉很抽象。把graph里的node/edges转成vector？然后用vector去train？

Graph 表示出来是一个N by N的sparse matrix。  N是 total number of nodes.  每个row理论上就是一个node的vector （但是是个sparse high dim vector/representation).

Network embeddings(eg DeepWalk)本质是做 dimension reduction, 把一个 1-by-N vector 变成1-by-d low dimension vector, 并且d-dim vectors preserve row/node-wise distance/proximity.

GNN (eg: Sage) 偏向deep learning art，学理上摸不清，但效果好。特点是可以用node本身的attribute当成feature然后用graph structure做feature aggregation.