Fix #10632: RandomNodeSplit fails on HeteroData

[JiwaniZakir]

Closes #10632

Before

RandomNodeSplit.__call__ in torch_geometric/transforms/random_node_split.py skipped any node store that lacked self.key (default: 'y'), regardless of the split mode. On HeteroData, this caused split='train_rest' to silently skip every node type that had no y attribute, producing a transformed object with no masks at all. Accessing .train_mask on the result then raised AttributeError: 'HeteroData' has no attribute 'train_mask'.

After

The skip condition in random_node_split.py (line 77) now gates on self.split != 'train_rest' in addition to the key check:

if (self.split != 'train_rest' and self.key is not None
        and not hasattr(store, self.key)):
    continue

train_rest distributes all remaining nodes to the training set without needing class labels, so it correctly processes every node store in a HeteroData object — including those without a y attribute (e.g., author nodes in MovieLens1M or OGB_MAG). test_rest and random still require the key to be present because they rely on per-class counts.

Changes

• torch_geometric/transforms/random_node_split.py: Added self.split != 'train_rest' guard to the early-continue condition so unlabeled node stores are not skipped when the split mode doesn't require y.
• test/transforms/test_random_node_split.py: Rewrote test_random_node_split_on_hetero_data to:
  • Increase author nodes from 300 → 1000 so num_val=100, num_test=200 fits within the store.
  • Explicitly test train_rest on HeteroData with both a labeled (paper) and unlabeled (author) node type, asserting exact mask cardinalities.
  • Add a second sub-test for test_rest confirming that author (no y) is still skipped while paper (has y) receives masks.

Testing

The updated test_random_node_split_on_hetero_data asserts:

data['author']: 4 attrs  → x, train_mask, val_mask, test_mask
data['paper']:  5 attrs  → x, y, train_mask, val_mask, test_mask

paper.train_mask.sum() == 1700   # 2000 - 100 - 200
paper.val_mask.sum()   == 100
paper.test_mask.sum()  == 200

author.train_mask.sum() == 700   # 1000 - 100 - 200
author.val_mask.sum()   == 100
author.test_mask.sum()  == 200

For test_rest with num_train_per_class=10, num_val=100, author (no y) is skipped (len == 1) and paper receives all three masks (len == 5).

---

This PR was created with AI assistance (Claude). The changes were reviewed by quality gates and a critic model before submission.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.18%. Comparing base (c211214) to head (623ce43).
⚠️ Report is 191 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10654      +/-   ##
==========================================
- Coverage   86.11%   84.18%   -1.93%     
==========================================
  Files         496      510      +14     
  Lines       33655    36019    +2364     
==========================================
+ Hits        28981    30323    +1342     
- Misses       4674     5696    +1022     

☔ 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.