Add LinkEncoder of GraphMixer
#7459
Conversation
akihironitta
force-pushed
the
feat/graph-mixer-link-encoder
branch
from
05280a8
to
3fb13a2
Compare
Codecov Report
@@ Coverage Diff @@
## master #7459 +/- ##
==========================================
- Coverage 91.89% 91.59% -0.31%
==========================================
Files 452 452
Lines 25292 25360 +68
==========================================
- Hits 23243 23229 -14
- Misses 2049 2131 +82
... and 17 files with indirect coverage changes 📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
Assuming each input node has the same number of temporal edges. If the number varies, the input needs zero-padding in advance to forward pass.
akihironitta
force-pushed
the
feat/graph-mixer-link-encoder
branch
from
22e05e2
to
8006e2b
Compare
torch_geometric/nn/encoding.py
Outdated
| edge_attr_time, _ = to_dense_batch( | ||
| edge_attr_time, | ||
| edge_index[1], | ||
| max_num_nodes=self.K, | ||
| batch_size=num_nodes, |
There was a problem hiding this comment.
That's a good point. If a destination node has degree larger than K, we will need to include K mst recent edges, i.e., largest K elements based on edge_time for the destination node.
|
I think this looks all good. Let's try to get this in. My only concern is that we should move this functionality to |
There was a problem hiding this comment.
@akihironitta Thanks for the great work.
@weihua916 I updated the PR, the latest K edge attributes are used to create the node embedding now. Appreciate review and approval from you before we merge.
Co-authored-by: Jintang Li <cnljt@outlook.com>
Co-authored-by: Jintang Li <cnljt@outlook.com>
This PR introduces the "link-encoder" defined in 3.1 of the GraphMixer paper, Do We Really Need Complicated Model Architectures For Temporal Networks?.