When making recommendation system, with each item for the node, the user's click stream generated goods and commodities between the sides, it generates a 3 million node size chart, but when doing node2vec, calculation of transition probability will spend thousands of hours, later changed to deepwalk method, modeling still takes a long time, has been running for 10 hours), it is only the beginning of modeling, haven't begun training vector generated word, I remember when node2vec method, see some people say that their company, especially the node, node2vec need seven hours running, why am I so slowly? My modeling process was wrong?

Deepwalk and node2vec code from the link below:
https://github.com/shenweichen/GraphEmbedding

Later I again with the authors of the paper presents the implementation of the methods to experimental deepwalk and node2vec, still found particularly slow,