Weighted Aggregations and Implementation of `SoftMedian`

[sigeisler]

As discussed in #6867 and #6871 I added a WeighedAggregation interface. For now, I provide 4 aggregations:

• WeightedMeanAggregation
• WeighedQuantileAggregation
• WeightedMedianAggregation
• SoftMedianAggregation (from paper Robustness of Graph Neural Networks at Scale)

I added an example with WeightedMeanAggregation and SoftMedianAggregation aggregations.

Is this how you @rusty1s imagined the extra interface? I am not sure if I like the distinction. It would be much nicer if there was a class inheritance, e.g., if Aggregation was the superclass of WeightedAggregation. However, this comes with some issues since the interfaces do not align. This is particularly an issue for using super().__call__.

Once, we agree on the architecture and implementation details, I will finish the documentation and add some unit tests.

PS: I also added the topk sparsification for GDC - of course I could also move this to an extra pull request if desired.

[sigeisler]

@rusty1s I know you are quite busy. However, a kindly reminder that I am still waiting for your feedback on how to implement the WeightedAggregation. Particularly, what class hierarchy would you prefer between Aggregation and WeightedAggregation?

[rusty1s]

Yeah, sorry for the delay here. I am on vacation this week and will try to get it in early next week.

[rusty1s]

Hi @sigeisler, made some changes and let WeightedAggregation inherit from Aggregation. Can you add some unit tests for the newly introduced weighted/soft median aggregations? Besides that, the PR seems ready to go.

[rusty1s]

For now, I removed integration in MessagePassing - I think this is better handled separately to test it properly.

[codecov[bot]]

Codecov Report

Merging #7025 (561483f) into master (c5dca4b) will increase coverage by 0.01%. The diff coverage is 91.15%.

@@            Coverage Diff             @@
##           master    #7025      +/-   ##
==========================================
+ Coverage   88.17%   88.18%   +0.01%     
==========================================
  Files         471      472       +1     
  Lines       28274    28417     +143     
==========================================
+ Hits        24930    25060     +130     
- Misses       3344     3357      +13     

Files	Coverage Δ
torch_geometric/nn/aggr/__init__.py	100.00% <100.00%> (ø)
torch_geometric/nn/aggr/quantile.py	100.00% <100.00%> (ø)
torch_geometric/nn/aggr/base.py	94.87% <95.83%> (+0.32%)	:arrow_up:
torch_geometric/nn/aggr/basic.py	92.52% <83.33%> (-0.62%)	:arrow_down:
torch_geometric/transforms/gdc.py	90.41% <92.85%> (+0.22%)	:arrow_up:
torch_geometric/nn/aggr/soft_median.py	90.00% <90.00%> (ø)

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more

[sigeisler]

Thank you for taking the time. I am planning to continue after the NeurIPS deadline

[rusty1s]

Thank you. Happy to get this in :)

[sigeisler]

Hi @rusty1s @wsad1 @EdisonLeeeee I apologize for the massive delay from start to finish of this pull request.

I have just fixed the last small things. I think this can be reviewed now.

[sigeisler]

Hi @rusty1s @EdisonLeeeee @wsad1

I know you are super busy with the many things you handle. However, I quickly wanted to check regarding the status.

[rusty1s]

Will try to get this in this week :) Thanks for the updates (missed the mail).

[sigeisler]

@EdisonLeeeee sorry for the delay. Can we do it the other way round? I just added this sparsification as it was missing/possible since PyTorch has a stable sort. However, the other feature is the one I am actually interested in having here.