Node2Vec implementation not the same as authors' implementation
wendywangwwt opened this issue · comments
Hi -
I noticed some differences between your implementation and the authors' implementation (https://github.com/aditya-grover/node2vec). Mainly the differences are the hyper-parameters used in Word2Vec:
| param | authors' | karateclub |
|---|---|---|
| hs | - (gensim default is 0) | 1 |
| sg | 1 | - (gensim default is 0) |
| learning rate | - (gensim default is 0.025) | 0.05, configurable |
| window size | 10, configurable | 5, configurable |
| workers | 8, configurable | 4, configurable (this one doesn't matter much) |
It could happen that the gensim recently changed the default values (https://radimrehurek.com/gensim/models/word2vec.html#gensim.models.word2vec.Word2Vec) though, as I didn't trace back to the historical doc.
Another part is the support for weighted graphs. From karateclub's source code, my understanding is that the implemented node2vec method doesn't support edge weights, while the authors' implementation does.
Can you please shed some light on why these are changed? The different setting of hs and sg are most confusing to me, as they likely have a large impact on the resulting embeddings.
Thank you!
The default hyperparameters are different. The model is the same. Have a nice day!