explainability

Untitled

decomposition: decompose the prediction

gradient-based: gradient w.r.t the input graph to identify the importance of the input portions

surrogate: fit a model based on the input and output

generation-based: synthesize the crucial patterns customized for an input, or generate key structures to globally explain the behavior of model predictions

perturbation-based: remove the unimportant edges and nodes so that the final prediction remains unchanged under such perturbations

counterfactual-based: identify the minimal substructure of the input which would change the original prediction if removed

self-explainable: make predictions and generate the corresponding explanations during the inference time, simultaneously

FederatedScope-GNN: Towards a Unified, Comprehensive and Efficient Package for Federated Graph Learning (KDD 2022 best paper)

GNN的联邦学习框架，主要是弥补这个研究方向的空缺，GNN的联邦学习要传输的东西和正常的联邦学习不太一样，感觉主要的不同就是这里。

图数据集的切分和FGL算法都有支持。

论文总结了联邦图学习的几个不同点：

  1. 传的数据是异构的，client之间的交互行为丰富很多；
  2. 切分数据集的挑战很多；

论文在阿里的电商推荐上用了这个联邦图学习框架，有提升。