twitter-research / feature-propagation Goto Github PK

If the edge index somehow does not include the last row or column of the adjacency matrix, the torch.sparse.FloatTensor() may not keep the full dimensionality of the matrix which later will result in error.
Change line 39 in feature_propagation.py
adj = torch.sparse.FloatTensor(edge_index, values=edge_weight).to(edge_index.device)
into
adj = torch.sparse.FloatTensor(edge_index, edge_weight,(n_nodes, n_nodes)).to(edge_index.device)

In the method get_propagation_matrix of the class FeaturePropagation the variable edge_weight is overriden in the following code:

edge_weight = edge_weight if edge_weight else torch.ones(edge_index.shape[1]).to(edge_index.device)
edge_index, edge_weight = get_symmetrically_normalized_adjacency(edge_index, n_nodes=n_nodes)

Since in the function get_symmetrically_normalized_adjacency all weights are defined as 1, this code essentially makes the graph unweighted.

I read your amazing paper “ On the Unreasonable Effectiveness of Feature Propagation in Learning on Graphs with Missing Node Features ”。 But I have questions about the results of Label Prop on Citeseer： I tuned hyperparameters but can't achieve 64.6 (I can only achieve 45 or so). Could you share the code or tell me where the results are from? Also, What is the Pos Enc method.

How should the FeaturePropagation algorithm be modified to apply to the directed graph?