Merge branch 'master' into tagdataset/add-llm-exp-pred

CHANGELOG.md

@@ -8,6 +8,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 ### Added
 
 - Added llm generated explanations to `TAGDataset` ([#9918](https://github.com/pyg-team/pytorch_geometric/pull/9918))
+- Added `Personalization` metric for link prediction ([#10015](https://github.com/pyg-team/pytorch_geometric/pull/10015))
 - Added `HitRatio` metric for link prediction ([#10013](https://github.com/pyg-team/pytorch_geometric/pull/10013))
 - Added Data Splitting Tutorial ([#8366](https://github.com/pyg-team/pytorch_geometric/pull/8366))
 - Added `Diversity` metric for link prediction ([#10009](https://github.com/pyg-team/pytorch_geometric/pull/10009))

test/metrics/test_link_pred_metric.py

@@ -12,9 +12,11 @@
     LinkPredMetricCollection,
     LinkPredMRR,
     LinkPredNDCG,
+    LinkPredPersonalization,
     LinkPredPrecision,
     LinkPredRecall,
 )
+from torch_geometric.testing import withCUDA
 
 
 @pytest.mark.parametrize('num_src_nodes', [100])
@@ -235,6 +237,28 @@ def test_diversity():
     assert pytest.approx(float(result)) == (1 + 2 / 3) / 2
 
 
+@withCUDA
+def test_personalization(device):
+    pred_index_mat = torch.tensor([[0, 1, 2, 3], [2, 1, 0, 4], [1, 0, 2, 5]],
+                                  device=device)
+    edge_label_index = torch.empty(2, 0, dtype=torch.long, device=device)
+
+    metric = LinkPredPersonalization(k=4).to(device)
+    assert str(metric) == 'LinkPredPersonalization(k=4)'
+    metric.update(pred_index_mat, edge_label_index)
+    result = metric.compute()
+    assert result.device == device
+    assert float(result) == 0.25
+    metric.reset()
+    assert metric.preds == []
+
+    metric.update(pred_index_mat[:0], edge_label_index)
+    result = metric.compute()
+    assert result.device == device
+    assert float(result) == 0.0
+    metric.reset()
+
+
 @pytest.mark.parametrize('num_src_nodes', [10])
 @pytest.mark.parametrize('num_dst_nodes', [50])
 @pytest.mark.parametrize('num_edges', [200])

torch_geometric/metrics/__init__.py

@@ -12,6 +12,7 @@
     LinkPredHitRatio,
     LinkPredCoverage,
     LinkPredDiversity,
+    LinkPredPersonalization,
 )
 
 link_pred_metrics = [
@@ -26,6 +27,7 @@
     'LinkPredHitRatio',
     'LinkPredCoverage',
     'LinkPredDiversity',
+    'LinkPredPersonalization',
 ]
 
 __all__ = link_pred_metrics

torch_geometric/metrics/link_pred.py

@@ -704,3 +704,86 @@ def compute(self) -> Tensor:
     def _reset(self) -> None:
         self.accum.zero_()
         self.total.zero_()
+
+
+class LinkPredPersonalization(_LinkPredMetric):
+    r"""A link prediction metric to compute the Personalization @ :math:`k`,
+    *i.e.* the dissimilarity of recommendations across different users.
+
+    Higher personalization suggests that the model tailors recommendations to
+    individual user preferences rather than providing generic results.
+
+    Dissimilarity is defined by the average inverse cosine similarity between
+    users' lists of recommendations.
+
+    Args:
+        k (int): The number of top-:math:`k` predictions to evaluate against.
+        batch_size (int, optional): The batch size to determine how many pairs
+            of user recommendations should be processed at once.
+            (default: :obj:`2**16`)
+    """
+    def __init__(self, k: int, batch_size: int = 2**16) -> None:
+        super().__init__(k)
+        self.batch_size = batch_size
+
+        if WITH_TORCHMETRICS:
+            self.add_state("preds", default=[], dist_reduce_fx="cat")
+            self.add_state('dev_tensor', torch.empty(0), dist_reduce_fx='sum')
+        else:
+            self.preds: List[Tensor] = []
+            self.register_buffer('dev_tensor', torch.empty(0))
+
+    def update(
+        self,
+        pred_index_mat: Tensor,
+        edge_label_index: Union[Tensor, Tuple[Tensor, Tensor]],
+        edge_label_weight: Optional[Tensor] = None,
+    ) -> None:
+        # NOTE Move to CPU to avoid memory blowup.
+        self.preds.append(pred_index_mat[:, :self.k].cpu())
+
+    def compute(self) -> Tensor:
+        device = self.dev_tensor.device
+        score = torch.tensor(0.0, device=device)
+        total = torch.tensor(0, device=device)
+
+        if len(self.preds) == 0:
+            return score
+
+        pred = torch.cat(self.preds, dim=0)
+
+        if pred.size(0) == 0:
+            return score
+
+        # Calculate all pairs of nodes (e.g., triu_indices with offset=1).
+        # NOTE We do this in chunks to avoid memory blow-up, which leads to a
+        # more efficient but trickier implementation.
+        num_pairs = (pred.size(0) * (pred.size(0) - 1)) // 2
+        offset = torch.arange(pred.size(0) - 1, 0, -1, device=device)
+        rowptr = cumsum(offset)
+        for start in range(0, num_pairs, self.batch_size):
+            end = min(start + self.batch_size, num_pairs)
+            idx = torch.arange(start, end, device=device)
+
+            # Find the corresponding row:
+            row = torch.searchsorted(rowptr, idx, right=True) - 1
+            # Find the corresponding column:
+            col = idx - rowptr[row] + (pred.size(0) - offset[row])
+
+            left = pred[row.cpu()].to(device)
+            right = pred[col.cpu()].to(device)
+
+            # Use offset to work around applying `isin` along a specific dim:
+            i = max(left.max(), right.max()) + 1  # type: ignore
+            i = torch.arange(0, i * row.size(0), i, device=device).view(-1, 1)
+            isin = torch.isin(left + i, right + i)
+
+            # Compute personalization via average inverse cosine similarity:
+            cos = isin.sum(dim=-1) / pred.size(1)
+            score += (1 - cos).sum()
+            total += cos.numel()
+
+        return score / total
+
+    def _reset(self) -> None:
+        self.preds = []