Variance in resulting embeddings about graph2vec HOT 4 CLOSED

Because of the transductive nature of implicit factorization embeddings from two separate runs are not comparable. You have different initializations and a non-convex optimization problem on which you run SGD. Results would be comparable if: You would have the same initialization, same order of SGD updates, same noise samples drawn.

from graph2vec.

Look at Appendix D of https://arxiv.org/pdf/1706.02216.pdf.

from graph2vec.

Because of the transductive nature of implicit factorization embeddings from two separate runs are not comparable. You have different initializations and a non-convex optimization problem on which you run SGD. Results would be comparable if: You would have the same initialization, same order of SGD updates, same noise samples drawn.

Is there any way to enforce the same initialization and other conditions in order to achieve comparable results?

I am asking as I am using Graph2Vec to process some data for a machine learning project. I am a bit concerned about the variations in the results.

from graph2vec.

from graph2vec.