Closed jasperzhong closed 9 months ago
PlatoGL用的模型叫MvDGAE,也是一个heterogeneous GNN. 感觉和HAN #328 差不多,也是每个metapath-based (second-order proximity)有一个graph encoder,然后把每个relation的output embedding aggregate. 不过它为了解决冷启动问题,多了个decoder,随机drop掉一些user-item然后做预测算一个loss加进去一起训练.
里面有一个sampling过程. 比如上图有UUI, UIU两个metapath,每个hop都sample若干个neighbors. 在PlatoGL里面每个hop sample 50个neighbors.
PlatoGL里面用到的metapath很多:
http://184pc128.csie.ntnu.edu.tw/presentation/21-11-09/Multi-view%20Denoising%20Graph%20Auto-Encoders%20on%20Heterogeneous.pdf