What does node2vec expect in .wv method?

[chuymtz]

Hello,

I;m working through chapter 4. In the prediction of the test set, you use train_mask_str in the book but in your gh code you use train_mask. Depending on which one you use one gets very different results.

if you pass the integer mask you get something that won't go pass 60% accuracy.

y_pred = clf.predict(node2vec.wv[test_mask])
acc = accuracy_score(y_pred, labels[test_mask])
print(f'Node2Vec accuracy = {acc*100:.2f}%')
>>>Node2Vec accuracy = 59.09%

But if i pass the string mask i get what you report both in gh and the book . Which seems really high considering how many samples you pick out in the train and test.

y_pred = clf.predict(node2vec.wv[test_mask_str])
acc = accuracy_score(y_pred, labels[test_mask])
print(f'Node2Vec accuracy = {acc*100:.2f}%')
>>>Node2Vec accuracy = 95.45%

How can one explain this discrepancy?

[mlabonne]

Hi, sorry for the very late reply! Thank you for noticing this discrepancy between the book and the code. Unfortunately, I can't replicate your issue. The data type should not matter here because it's masking the same nodes.

[chuymtz]

No problem. thanks you for the reply (and amazing book). I think this is more of an issue with node2vec. I'll take what I learned conceptually and move on 😊 thanks!

[wjn-1511]

I've figured it out. The issue was how the embeddings were indexed. I'd like to explain this by comparing Deep Walk and Node2Vec.

In DeepWalk code, we manually created a numpy array nodes_wv where the embeddings were explicitly sorted by node ID (nodes_wv[i] was the vector for node i). This allowed us to correctly use an integer mask for slicing both the features and the labels.

However, in Node2Vec code, we directly use the gensim wv object. Its internal order is not based on node ID, so indexing it with integers caused a misalignment between the features and the labels. This explains the discrepancy in performance.