search

YuxiangRen / Heterogeneous-Deep-Graph-Infomax

HDGI code
59 stars 14 forks source link

other experiments #8

Open xinchen1412 opened 4 years ago

xinchen1412 commented 4 years ago

Great respect for your work. Have you conducted other experiments? I used TSNE for visualization and found that the results were not ideal. I don't know if this is normal. Looking forward to your answer. I hope I didn't disturb you.

YuxiangRen commented 4 years ago

I didn't use TSNE for visualization in my experiments. I just use node classification and clustering to evaluate the representations. I am not sure whether the dimensionality reduction of TSEN will affect the learned representations, but it's a quite interesting observation. I may need to spend some time thinking about that.

If you have more observations or more questions, please let me know.

xinchen1412 commented 4 years ago

Thank you for your reply! I'll try your suggestions. One more question: Is the NMI value in the code the same as the NMI value in the paper? I used the "ACM" and the "hdgi-c"model, I got an average NMI value of 0.49 and an average ARI value of 0.46. The values in paper are about 50. Do you have any suggestions?

YuxiangRen commented 4 years ago

The way I calculate NMI and ARI for HDGI-C is the same as HDGI-A. It is implemented as the code in DGI-HGAT/utils/clustering.py. You can check it. Your reported values are something weird because they are too low for the ACM dataset. I have no idea about these values, but you can check my code.

xinchen1412 commented 4 years ago

I used the code in "DGI-HGAT/utils/clustering.py" to calculate NMI and ARI, but the clustering results are so strange, the classification result is normal(about 0.9). I added the following code to "execute.py":

labels_np = labels_np.to(torch.device("cpu")).numpy()
clustering.my_Kmeans(embeddings, labels_np, k=3, time=10, return_NMI=False)
YuxiangRen commented 4 years ago

You may need to check whether the labels_np is constructed correctly. You can refer to executeCla.py in DGI-HGAT.

Fujiaoji commented 4 years ago

Thank you for your reply! I'll try your suggestions. One more question: Is the NMI value in the code the same as the NMI value in the paper? I used the "ACM" and the "hdgi-c"model, I got an average NMI value of 0.49 and an average ARI value of 0.46. The values in paper are about 50. Do you have any suggestions?

I used IMDB data set and the result of NMI is 0.0265, while 0.6324 in DBLP. Did u figure out?