Add LPFormer model and example

[HarryShomer]

This pull request adds an implementation of the SOTA Link Prediction model LPFormer from the KDD'24 paper - LPFormer: An Adaptive Graph Transformer for Link Prediction.

Example

python examples/lpformer.py.

I am able to replicate the reported performance on ogbl-ppa.

The results are shown below. I use the same hyperparameters used in the paper. Results are over 5 seeds.

Reported	PyG
63.32 ± 0.6	63.15 ± 0.9

Testing

The tests for LPFormer can be found in test/nn/models/test_lpformer.py. They each pass.

Docs

I've also included the documentation for the code. It builds and renders correctly with Sphinx with no errors.

[codecov]

Codecov Report

❌ Patch coverage is 88.95899% with 35 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.93%. Comparing base (c211214) to head (ff86503).
⚠️ Report is 138 commits behind head on master.

Files with missing lines	Patch %	Lines
torch_geometric/nn/models/lpformer.py	88.92%	35 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #9956      +/-   ##
==========================================
- Coverage   86.11%   85.93%   -0.18%     
==========================================
  Files         496      502       +6     
  Lines       33655    35128    +1473     
==========================================
+ Hits        28981    30189    +1208     
- Misses       4674     4939     +265     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[HarryShomer]

@wsad1 @EdisonLeeeee Can I get an ETA on a review? Thanks!

[puririshi98]

can you update this to use modern pyg (torch-geometric and pyg-lib) not deprecated packages such as torch_sparse

[HarryShomer]

@puririshi98 Thanks for the heads up!

I'm curious though, what's the alternative to torch_sparse? As far as I can tell, pyg still doesn't support torch.sparse for memory-efficient aggregations, which is why I use SparseTensor.

Could you let me know what I should use instead of torch_sparse?

[puririshi98]

@puririshi98 Thanks for the heads up!

I'm curious though, what's the alternative to torch_sparse? As far as I can tell, pyg still doesn't support torch.sparse for memory-efficient aggregations, which is why I use SparseTensor.

Could you let me know what I should use instead of torch_sparse?

there should be a pytorch native version, look into it, if you struggle to find an option let me know and i can try to check

[HarryShomer]

@puririshi98 I've updated the code to remove the torch_sparse references.

For some reason I was under the impression that the MessagePassing layers could not accept torch.sparse for the edge_index argument. I think this may have been true for some older versions.

Please let me know if there's anything else needed on my side.

[puririshi98]

will merge once we get the onnx issue fixed upstream and merge that in . aiming EoD tmrw

[puririshi98]

LGTM now