Closed jasperzhong closed 1 year ago
很可惜,这篇temporal heterogeneous graph定义看上去是CTDG,实际上还是DTDG. 这篇paper方法比较奇葩,全程都是用influence of node pairs (两个node embedding的L2 norm的负数)带进去作attention各种操作...
metapath before time t定义成all of temporal edges it included generated before time t,然后优先选择比较靠近给定的time t的n个metapath instances. 最后实验就是DTDG,分成T个timestamp. 粒度还是DTDG.
THINE 效果击败了 MAGNN #340 和 DyHNE #334
https://www.ijcai.org/proceedings/2021/0203.pdf