Question about the deepwalk implementation about graphsage HOT 4 CLOSED

Hi!

This is a great question: We actually use larger window sizes, but we do the random walks as a pre-processing step. In particular, for the unsupervised version you need to specify a file "<train_prefix>-walks.txt" which gives pairs of nodes that co-occur on random walks. You can specify different "window sizes" in how you generate this file. As mentioned in our experiment, we took random walks of length 5, meaning that we used a "window size" of 10 (if that makes sense).

Cheers,
Will

from graphsage.

Got it. Thanks for the explanation.

-Minjie

from graphsage.

@williamleif Just wondering why you did not use the deepwalk implementation by the authors for comparison? Did you see any difference in results?

from graphsage.

Hi!

This is a great question: We actually use larger window sizes, but we do the random walks as a pre-processing step. In particular, for the unsupervised version you need to specify a file "<train_prefix>-walks.txt" which gives pairs of nodes that co-occur on random walks. You can specify different "window sizes" in how you generate this file. As mentioned in our experiment, we took random walks of length 5, meaning that we used a "window size" of 10 (if that makes sense).

Cheers,
Will

Hi Will, it seems like the run_random_walks function in the code only uses window_size 1, also it is hard-coded.

from graphsage.