Adding TGCN for nodeproppred by ntgbaoo · Pull Request #89 · tgm-team/tgm

ntgbaoo · 2025-07-03T03:31:10Z

Objective

Adding the implementation of T-GCN and a node property prediction example for T-GCN

New scripts

Testing evidence

Command used:

python .\examples\nodeproppred\tgcn.py --dataset tgbn-trade --time-gran Y --batch-time-gran Y --epochs 100 --seed 720

python .\examples\nodeproppred\gclstm.py --dataset tgbn-trade --time-gran Y --batch-time-gran Y --epochs 100 --seed 720

Cross Entropy Loss & validation NDCG over epochs:

Achieved test result (no early stopping): 0.2521 NDCG on tgbn-trade (GCLSTM achieved 0.2462 NDCG)

TODO for another PR:

Add link prediction example for TGCN

Added initial development of TGCN and updated __init__.py with new model class

Fixed importing wrong lib for zeros() init. Added NodePropPred example for TGCN. Added Integration test for TGCN

ntgbaoo · 2025-07-03T03:44:57Z

Not sure why the test check on GitHub failed. The test ran successfully on my local (except 3 tests failed due to the issue mentioned in this PR #82)

Jacob-Chmura · 2025-07-03T10:24:32Z

Not sure why the test check on GitHub failed

Safe to ignore these tests. There was an issue in the test cases due to auto-merging conflicts in some PRs that I pushed. I have adressed these and the fix should be in master shortly.

Jacob-Chmura

Very nice work. Just some minor comments. I am not familiar with the TGCN paper, @shenyangHuang please review the core library layer and make a judgement on the expected performance.

Jacob-Chmura · 2025-07-03T14:30:13Z

+        '--gres=gpu:a100l:1',
+    ]
+)
+def test_gclstm_nodeprop_pred(slurm_job_runner, dataset):


Suggested change

def test_gclstm_nodeprop_pred(slurm_job_runner, dataset):

def test_tgcn_nodeprop_pred(slurm_job_runner, dataset):

Jacob-Chmura · 2025-07-03T14:30:53Z

+r"""python -u gclstm.py --dataset tgbn-trade --time-gran r --batch-time-gran r
+python -u gclstm.py --dataset tgbn-genre --time-gran s --batch-time-gran D\
+example commands to run this script.
+"""


Suggested change

r"""python -u gclstm.py --dataset tgbn-trade --time-gran r --batch-time-gran r

python -u gclstm.py --dataset tgbn-genre --time-gran s --batch-time-gran D\

example commands to run this script.

"""

"""

I think this can be removed

Jacob-Chmura · 2025-07-03T14:34:46Z

+        edge_index = torch.stack([batch.src, batch.dst], dim=0)
+        edge_weight = batch.edge_weight if hasattr(batch, 'edge_weight') else None  # type: ignore
+        z, h_0 = self.encoder(node_feat, edge_index, edge_weight, h_0)
+        z_node = z[batch.nid_to_idx[batch.node_ids]]  # type: ignore


After merging #60, the batch attribute to use is actually global_to_local.

Suggested change

z_node = z[batch.nid_to_idx[batch.node_ids]] # type: ignore

z_node = z[batch.global_to_local[batch.node_ids]] # type: ignore

Jacob-Chmura · 2025-07-03T14:35:56Z

+        self,
+        x: torch.Tensor,
+        edge_index: torch.Tensor,
+        edge_weight: torch.Tensor,


Suggested change

edge_weight: torch.Tensor,

edge_weight: torch.Tensor | None,

Jacob-Chmura · 2025-07-03T14:37:37Z

+parser.add_argument('--seed', type=int, default=1337, help='random seed to use')
+parser.add_argument('--dataset', type=str, default='tgbn-genre', help='Dataset name')
+parser.add_argument('--device', type=str, default='cpu', help='torch device')
+parser.add_argument('--epochs', type=int, default=100, help='number of epochs')


Just double checking if we want to use 100 or 10 as default number of epochs

I would say 100 epochs

for DTDG methods usually takes more epochs

Jacob-Chmura · 2025-07-03T14:38:59Z

+            edge_index (PyTorch Long Tensor): Graph edge indices.
+            edge_weight (PyTorch Long Tensor, optional): Edge weight vector.
+            H (PyTorch Tensor, optional): Hidden state matrix for all nodes.
+            C (PyTorch Tensor, optional): Cell state matrix for all nodes.


Suggested change

C (PyTorch Tensor, optional): Cell state matrix for all nodes.

Unless we need this tensor in which case example should be updated

codecov · 2025-07-05T01:22:12Z

Codecov Report

Attention: Patch coverage is 29.62963% with 38 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
tgm/nn/recurrent/tgcn.py	26.92%	38 Missing ⚠️

📢 Thoughts on this report? Let us know!

shenyangHuang · 2025-07-05T12:30:36Z

test performance of 0.3654 is quite sensible.
for tgbn-trade TGN achieves 0.374 test. So I would say the performance is quite reasonable.
see here for the leaderboard.

shenyangHuang

performance looks good.

shenyangHuang · 2025-07-05T12:32:03Z

+        in_channels: int,
+        out_channels: int,
+        improved: bool = False,
+        cached: bool = False,


when would you want to turn this one? what would be the benefit?

improved and cached are both used to configure GCNConv. By default, they are set to False. From what I am understanding from reading the documentation from PyG.

If we set improved = True , the self-loops are added A+2I instead of A+I. This increases the weight of self-loops, giving each node’s own features more influence during aggregation.

If we set cached = True, the layer computes the normalized adjacency matrix only once, and reuses it. This speeds up training and inference but is only suitable for transductive learning. Scenarios where cached can be set to True include tasks such as traffic forecasting, where graph structure is assumed to be static and only node features are dynamic. cached should not be set to False when graph structure dynamically changes.

Thanks for pointing this out. These are all useful information that we should include in the comment section.

yeah thanks for adding the comments here, much more clear now

shenyangHuang · 2025-07-05T12:34:14Z

+parser.add_argument('--seed', type=int, default=1337, help='random seed to use')
+parser.add_argument('--dataset', type=str, default='tgbn-genre', help='Dataset name')
+parser.add_argument('--device', type=str, default='cpu', help='torch device')
+parser.add_argument('--epochs', type=int, default=100, help='number of epochs')


for DTDG methods usually takes more epochs

Increase default number of epoch to 250

ntgbaoo · 2025-07-06T01:07:56Z

I added further information about cached and improved in the comment section and increased the default number of epochs as you suggested, @shenyangHuang. Please take a look whenever you have a chance.

shenyangHuang · 2025-07-06T11:46:38Z

+        in_channels: int,
+        out_channels: int,
+        improved: bool = False,
+        cached: bool = False,


yeah thanks for adding the comments here, much more clear now

shenyangHuang · 2025-07-11T09:45:50Z

The new performance makes sense to me, I got similar range from GCN yesterday for nodeproppred.

Bảo Ngô added 2 commits July 2, 2025 19:36

Initial development of TGCN

4d3a8e1

Added initial development of TGCN and updated __init__.py with new model class

Finalized & Tested TGCN implementation

427ff60

Fixed importing wrong lib for zeros() init. Added NodePropPred example for TGCN. Added Integration test for TGCN

ntgbaoo requested review from Jacob-Chmura and shenyangHuang July 3, 2025 03:31

Jacob-Chmura requested changes Jul 3, 2025

View reviewed changes

Jacob-Chmura mentioned this pull request Jul 3, 2025

Adding MRR for existing models #37

Closed

2 tasks

ntgbaoo self-assigned this Jul 4, 2025

ntgbaoo changed the title ~~Adding GCLSTM for nodeproppred~~ Adding TGCN for nodeproppred Jul 5, 2025

correct typos and remove redundant comment as Jacob suggested

53d65ba

shenyangHuang reviewed Jul 5, 2025

View reviewed changes

Added additional information about cached and improved input to TGCN.

d9accd0

Increase default number of epoch to 250

ntgbaoo requested a review from shenyangHuang July 6, 2025 01:05

shenyangHuang approved these changes Jul 6, 2025

View reviewed changes

ntgbaoo requested a review from Jacob-Chmura July 6, 2025 17:40

Bảo Ngô added 2 commits July 10, 2025 19:53

Merge branch 'main' into 79-tgcn-nodePropPrediction

e7f541f

Test TGCN with --time-gran Y --batch-time-gran Y

79b6203

ntgbaoo requested a review from shenyangHuang July 11, 2025 02:27

Jacob-Chmura approved these changes Jul 11, 2025

View reviewed changes

shenyangHuang approved these changes Jul 11, 2025

View reviewed changes

ntgbaoo merged commit cdf618d into main Jul 11, 2025
6 checks passed

ntgbaoo deleted the 79-tgcn-nodePropPrediction branch July 12, 2025 02:01

	def test_gclstm_nodeprop_pred(slurm_job_runner, dataset):
	def test_tgcn_nodeprop_pred(slurm_job_runner, dataset):

	z_node = z[batch.nid_to_idx[batch.node_ids]] # type: ignore
	z_node = z[batch.global_to_local[batch.node_ids]] # type: ignore

Uh oh!

Conversation

ntgbaoo commented Jul 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Objective

New scripts

Testing evidence

Command used:

TODO for another PR:

Uh oh!

ntgbaoo commented Jul 3, 2025

Uh oh!

Jacob-Chmura commented Jul 3, 2025

Uh oh!

Jacob-Chmura left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov Bot commented Jul 5, 2025

Codecov Report

Uh oh!

shenyangHuang commented Jul 5, 2025

Uh oh!

shenyangHuang left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ntgbaoo Jul 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ntgbaoo commented Jul 6, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shenyangHuang commented Jul 11, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ntgbaoo commented Jul 3, 2025 •

edited

Loading

shenyangHuang left a comment •

edited

Loading

ntgbaoo Jul 6, 2025 •

edited

Loading