Add support for projecting features before aggregation in SAGEConv

[wilcoln]

This pull request adds the implementation for Max and Mean Pool aggregations for SAGEConv which were implemented in the original study Inductive Representation Learning on Large Graphs, it's a sequel to PR #4379, which in turn is related to issue #1147.

The implementation is a direct application of the Eq 3 of the original paper (for both mean and max):

cc @hunarbatra

[codecov]

Codecov Report

Merging #4437 (eca3e49) into master (3f0019f) will decrease coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4437      +/-   ##
==========================================
- Coverage   82.69%   82.63%   -0.06%     
==========================================
  Files         313      313              
  Lines       16240    16259      +19     
==========================================
+ Hits        13429    13435       +6     
- Misses       2811     2824      +13     

Impacted Files	Coverage Δ
torch_geometric/nn/conv/sage_conv.py	100.00% <100.00%> (ø)
torch_geometric/data/hetero_data.py	93.10% <0.00%> (-2.30%)	⬇️
torch_geometric/data/data.py	90.70% <0.00%> (-1.41%)	⬇️
torch_geometric/loader/link_neighbor_loader.py	89.52% <0.00%> (-0.80%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3f0019f...eca3e49. Read the comment docs.

[rusty1s]

Thank you for the PR. I am not sure I fully understand this. How does that differ from aggr=mean and aggr=max? IMO, these already implement the linked Equation (we just added LSTM support to make SAGEConv finally support all proposed aggregations).

[wilcoln]

Thanks for your comment @rusty1s .

As you know, the mean/max aggregator directly applies the mean/max reduction to all neighboring nodes.

While the mean/max pool aggregator first applies an additional linear layer (+ bias) followed by an activation function, and only then, performs the element-wise mean/max reduction.

[rusty1s]

Thanks for the clarification. How about we add an option to apply the transformation before aggregation rather than after it? This way, there would be no need to modify the underlying aggregation scheme.

[wilcoln]

That would be much cleaner indeed and would work with all existing aggregators, I will update my PR accordingly, by adding a pool boolean parameter to the class constructor, defaulting to False. How does that sound?

[rusty1s]

Sounds good to me, thank you very much! I am wondering if pool is the most descriptive name for this. How about we name it pre_transform or something similar?

[rusty1s]

Thank you!