Fix missing hyperedge weight application in HypergraphConv

test/nn/conv/test_hypergraph_conv.py

@@ -46,3 +46,27 @@ def test_hypergraph_conv_with_more_edges_than_nodes():
     assert out.size() == (num_nodes, out_channels)
     out = conv(x, hyperedge_index, hyperedge_weight)
     assert out.size() == (num_nodes, out_channels)
+
+
+def test_hypergraph_conv_weight_effect():
+    """Test that non-uniform hyperedge weights produce different results
+    from uniform weights, verifying that W is applied correctly per the
+    formula X' = D^{-1} H W B^{-1} H^T X Theta.
+    """
+    in_channels, out_channels = (16, 32)
+    hyperedge_index = torch.tensor([[0, 0, 1, 1, 2, 3], [0, 1, 0, 1, 0, 1]])
+    num_nodes = hyperedge_index[0].max().item() + 1
+    x = torch.randn((num_nodes, in_channels))
+
+    conv = HypergraphConv(in_channels, out_channels)
+
+    # Uniform weights (all ones) should match default (no weight argument)
+    uniform_weight = torch.ones(2)
+    out_default = conv(x, hyperedge_index)
+    out_uniform = conv(x, hyperedge_index, uniform_weight)
+    assert torch.allclose(out_default, out_uniform, atol=1e-6)
+
+    # Non-uniform weights should produce different results
+    nonuniform_weight = torch.tensor([1.0, 0.5])
+    out_nonuniform = conv(x, hyperedge_index, nonuniform_weight)
+    assert not torch.allclose(out_default, out_nonuniform, atol=1e-4)

torch_geometric/nn/conv/hypergraph_conv.py

@@ -191,6 +191,16 @@ def forward(self, x: Tensor, hyperedge_index: Tensor,
 
         out = self.propagate(hyperedge_index, x=x, norm=B, alpha=alpha,
                              size=(num_nodes, num_edges))
+
+        # Apply hyperedge weights W to the intermediate edge representation.
+        # Per the formula X' = D^{-1} H W B^{-1} H^T X Theta, the diagonal
+        # weight matrix W must scale edge features between the two message
+        # passing steps (node-to-edge followed by edge-to-node):
+        if self.use_attention:
+            out = out * hyperedge_weight.view(-1, 1, 1)
+        else:
+            out = out * hyperedge_weight.view(-1, 1)
+
         out = self.propagate(hyperedge_index.flip([0]), x=out, norm=D,
                              alpha=alpha, size=(num_edges, num_nodes))