Replicating the results in the paper about graphsage HOT 3 OPEN

Hi,

That is strange because according to my experiment logs, the default settings (learning rate 0.01 and a "small" model) should be best for graphsage-mean. Across five runs I got scores of 0.602, 0.592, 0.588, 0.602, and 0.607 on the test set, i.e. a mean of 0.598 and standard deviation of 0.007. Your number is slightly lower than I would expect for random variation given the low variance I saw in my trials, but there is of course the possibility of weird TensorFlow version differences etc. Are you referring to the test set performance or performance on the whole dataset (e.g., including the val set)?

Also, in more recent work https://arxiv.org/pdf/1710.10568.pdf achieve >0.95 F1 on the PPI dataset using a couple minor changes to GraphSAGE (e.g., larger model and adding layer norm), so it should definitely be possible to get higher numbers without too much hassle.

Anyways, sorry that you are having difficulties and let me know how it goes!

from graphsage.

The default hyper-parameters yield an average f1_micro of around 0.517.

Optimization Finished!
Full validation stats: loss= 0.51058 f1_micro= 0.51657 f1_macro= 0.31484 time= 0.11630
Writing test set stats to file (don't peak!)

I just do as the following:

1. clone this project.
2. Change the line "FROM tensorflow/tensorflow:1.3.0-gpu" in Dockerfile.gpu. (I also tried 1.5.0-gpu version)
3. build & run in docker with example_supervised.sh

from graphsage.

Hi,

In my case of experiment, the default parameters (lr=0.01, neighbor samples of 10 x 25, hidden layer dim=128) running with 10 epochs yielded micro F1 score in test set of 0.57455. While the same parameters running with 20 epochs got 0.5984.

from graphsage.