search

thunlp / NEU

source code for IJCAI 17 paper "Fast Network Embedding Enhancement via High Order Proximity Approximation"
MIT License
33 stars 10 forks source link

A question about the result of Multi-label Classification task on Cora dataset #1

Closed BenchengY closed 7 years ago

BenchengY commented 7 years ago

I wander that why the auc is so high on cora for Classification task in your paper. Do you use the text information of cora? As i run node2vec on cora(just use citation links information) for Classification the auc is about 30%. thanks!

albertyang33 commented 7 years ago

No, I just use the link information. The AUC value measures the probability that the score of a missing edge is larger than that of a nonexistent edge. Thus AUC is certainly larger than 50%. I think you should double check your evaluation code.

BenchengY commented 7 years ago

I know the auc is certainly larger than 50% for link prediction task. but what i want know is for node classification task on cora . thanks.