#3 GraphSAGE pseudocode/demo

[kjun9]

No description provided.

[youph]

Code Review

Comments

The code runs, the loss decreases during training as expected. However, the f1 scores do not seem to change along with the loss - this is somewhat suspicious, as I'd expect the f1 scores to increase while the loss decreases. Might be worth checking how the f1 scores are evaluated.
The code is lean and reasonably easy to follow, even without comments.

Suggestions for @kjun9

• Please add comments to help understand the code's purpose and workflow
• Perhaps add a sentence or two referencing the example use case in the paper, being done by this code?
• The final results on the test set are cryptic, could you please make them a bit more verbose, so that at least it was clear that it was the final test score?
• Might be worth checking how the f1 scores are evaluated (see the Comments above).

[kjun9]

• python -m example.example_graphsage should be what was in the README already, could you double check you're seeing this?
• I'll make the other changes and also double check the f1 scores from the original implementation to see if I've missed something.

[youph]

Yes, the run command was correct, my bad.

[kjun9]

@youph Thanks for the suggestions, please review the changes I've made accordingly. Most importantly, I found out I was missing relu's in between each of the aggregation layers (despite having drawn boxes for them for the presentation...), and so now the accuracy results look a lot more comparable to the original implementation. Also added some more comments and optional parameters and a little bit of info about the example data from the original graphSAGE codebase

[youph]

@kjun9 Thanks for fixing the code. Yes, now the f1 scores look better; in particular, they improve together with the loss.
May I ask you to also add description (purpose, input args, output) to
def create_iterator(graph, nl, nf): (line 24 in example_graphsage.py)

[kjun9]

@youph sure, you can check the descriptions I've added now

[youph]

@kjun9 Thanks, all good now.