search

thunlp / OpenNE

An Open-Source Package for Network Embedding (NE)
MIT License
1.68k stars 485 forks source link

Why not use the Classifier(LogisticRegression) to evaluate the performance of GCN? #60

Closed BenchengY closed 4 years ago

BenchengY commented 5 years ago

It seems that the acc of GCN is directly obtained from the GCN net, while the acc of other methods are obtained from the Classifier(LogisticRegression). Is it fair to compare the performance between GCN and others by different classifiers (i.e., GCN net and LogisticRegression)? Why not use the Classifier(LogisticRegression) to evaluate the performance of GCN?

zzy14 commented 5 years ago

GCN is a semi-supervised method while other methods are unsupervised network embedding methods. GCN can directly predict the labels of nodes.

BenchengY commented 5 years ago

Although GCN can directly predict the labels of nodes， is it fair to compare the performance between GCN and others by different classifiers (i.e., GCN net and LogisticRegression)?

Bznkxs commented 4 years ago

Hi BenchengY, GCN is used in semi-supervised node classification, which is a different task from unsupervised node classification (with methods such as Node2vec and GraRep). There is no way to compare these two tasks. I think it is also unfair to put GCN embeddings into a classifier and compare with other methods by classifiers.