Can HypergraphConv used for directed Hypergraph

[CaiYitao]

we have a chemical reaction network, which can be represented as a derivation graph/ directed hypergraph. I am trying to use HypergraphConv class to do convolution on it. But it seems not working.
Please see attached hypergraph. the r_i are hyperedges, molecules are nodes. and the hyperedges indicate reaction rules/types which connect reactants to products.

here is my code

from torch_geometric.nn import HypergraphConv

class HyperGraphLayer(nn.Module):
    def __init__(self, cfg):
        super().__init__()


        self.dim_h = cfg.dim_h
        self.num_heads = cfg.num_heads
        self.attn_dropout = cfg.attn_dropout
        self.layer_norm = cfg.layer_norm
        self.batch_norm = cfg.batch_norm
        # self.lin = nn.Linear(cfg.dim_h * cfg.hg_attn_heads, cfg.dim_h)
        self.hypergraph_conv = HypergraphConv(cfg.dim_h, cfg.dim_h, use_attention=True)
        self.attn = nn.MultiheadAttention(cfg.dim_h, cfg.num_heads, dropout=self.attn_dropout, batch_first=True)

        self.mlp = nn.Sequential(
            nn.Linear(cfg.dim_h, cfg.dim_h*2),
            nn.GELU(),
            nn.Dropout(cfg.dropout),
            nn.Linear(cfg.dim_h*2, cfg.dim_h),
            nn.GELU(),
            nn.Dropout(cfg.dropout)
        )

        self.layer_norm_local = nn.LayerNorm(cfg.dim_h)
        self.layer_norm_attn = nn.LayerNorm(cfg.dim_h)
        self.layer_norm = nn.LayerNorm(cfg.dim_h)
        
    def forward(self, hypergraph):
        h = hypergraph.x
        h_in = h

        print("hyperedge_index: ", hypergraph.edge_index, "hyperedge_attr shape: ", hypergraph.edge_attr.shape)
        print("hyperedge_index[1]", hypergraph.edge_index[1] )
     
        h_local = self.hypergraph_conv(h, hypergraph.edge_index, hyperedge_attr = hypergraph.edge_attr)
       
        h_local = h_local + h_in
        h_local = self.layer_norm_local(h_local)
        h_attn = self.attn(h, h, h, need_weights=False)[0]
   
        h_attn = h_attn + h_in
        h_attn = self.layer_norm_attn(h_attn)

        h= h_local + h_attn
        h = h + self.mlp(h)
        h = self.layer_norm(h)

        hypergraph.x = h
        return hypergraph

but it gives error:

hyperedge_index: tensor([[5, 1, 4, 3, 2, 6],
        [0, 0, 0, 1, 1, 1]]) 
hyperedge_index[1]: tensor([0, 0, 0, 1, 1, 1])
hyperedge_attr shape:  torch.Size([1, 128])
Traceback (most recent call last):
  
    h_local = self.hypergraph_conv(h, hypergraph.edge_index, hyperedge_attr = hypergraph.edge_attr)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
    x_j = hyperedge_attr[hyperedge_index[1]]
          ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
IndexError: index 1 is out of bounds for dimension 0 with size 1

since currently the hypergraph only have one hyperedge which from reactants set mol 5, mol1, mol4 to products set mol3, mol2, mol6. the hypergraph conv seems can not handle this case.
I am not sure if I understand correctly about your HypergraphConv document, I assume that hyperedge_index: tensor([[5, 1, 4, 3, 2, 6],
[0, 0, 0, 1, 1, 1]]) means that the nodes 5, 1 ,4 connected with 3 ,2 ,6 by a directed hyperedge.
Looking forward to your suggestions.

summary_2.pdf

[rusty1s]

Sorry for late reply. Why does your hyperedge_attr only holds features for a single hyperedge? Changing this should resolve your issue.

[CaiYitao]

thank you for your kind reply. in the printout case, the hypergraph now only have one hyperedge, our hypergraph is directed hypergraph. can the HypergraphConv used for directed one? from the paper in the doc suggested, it seems not the case. did I misunderstood anything?

[rusty1s]

What do you mean by directed hypergraph? Hypergraphs are always directed, right?

[CaiYitao]

It seems not the case, directed hypergraph have two incidence matrix one is source(tail) nodes to hyperedge and the other is hyperedge to target(head) nodes

I find the following papers are relevant.
You can refer to these papers.

IET Intelligent Trans Sys - 2021 - Luo - Directed hypergraph attention network for traffic forecasting.pdf
heterogenousHyperGraphNN.pdf

Directed_Hypergraph_Neural_Network.pdf

[CaiYitao]

Could you please explain the concept of hyperedge_index in the context of the HypergraphConv class? Here's my understanding:

For example, I believe hyperedge_index: tensor([[5, 1, 4, 3, 2, 6, 7, 9, 8, 11, 12], [0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1]]) represents directed hyperedges in a hypergraph. It seems to indicate the following:

• Hyperedge 1: Source nodes (5, 1, 4) are connected to target nodes (3, 2, 6).
• Hyperedge 2: Source nodes (7, 9) are connected to target nodes (8, 11, 12).

The 0s in the second row denote source nodes, and 1s denote target nodes. Transitions from 1 to 0 likely mark the beginning of new hyperedges.

Could you please confirm if this understanding is in the same page as in the HypergraphConv class?

[rusty1s]

hyperedge_index should represent the non-zero indices of the incidence matrix of a hypergraph. In your case,

hyperedge_index: tensor([[5, 1, 4, 3, 2, 6, 7, 9, 8, 11, 12], [0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1]])

this means that [5, 1, 4, 7, 9] define a hyperedge, and [3, 2, 6, 8, 11, 12] define a hyperedge (i.e., all indices in hyperedge_index[0] that share an index in hyperedge_index[1].

[CaiYitao]

Thank you very much! so if we need to construct a new directed hypergraph conv class, the hyperedge index of this example should be
hyperedge_index: tensor([[5, 1, 4, 3, 2, 6, 7, 9, 8, 11, 12], [0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1]]) and then we build an extra index indicate the source to target, idx_source_targe: tensor([0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1]) 0s indicate source/tail, 1s indicate target/head of each hyperedge, from 1 to 0 indicates another hyperedge.

the following is a real case

after I have done such changes,
hyperedge_index after transform: tensor([[5, 6, 7], [0, 0, 0]]) indicate there is only one hyperedge connect with 5,6,7. and idx_source_target: tensor([0, 1, 1]),
the node indices 5, 6, 7 came from a derivation graph which indexed by [0, numVertices + numEdges].
There is an error in index_select for pytorch geometric.

   edge_attr_i = self._index_select(edge_attr, edge_index_i)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "/lib/python3.11/site-packages/torch_geometric/nn/conv/message_passing.py", line 300, in _index_select
   return self._index_select_safe(src, index)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "/lib/python3.11/site-packages/torch_geometric/nn/conv/message_passing.py", line 315, in _index_select_safe
   raise IndexError(
IndexError: Found indices in 'edge_index' that are larger than 0 (got 7). Please ensure that all indices in 'edge_index' point to valid indices in the interval [0, 1) in your node feature matrix and try again.

if we have a hyperpergraph which does not have hyperedges ,only the individual vertices, at the beginning, and at some time point there are hyperedges appeared, so the vertices indices within hyperedges could happen to have [ 5, 6, 7 ] , or something like[10, 28, 62,...] that. Does it mean that we should reindex the vertices within hyperedges to be selfcontained in hyperedges? and I do not understand why we should ensure that all indices in 'edge_index' point to valid indices in the interval [0, 1) in your node feature matrix

[CaiYitao]

It is actually related to this issue: qiankunmu/HDHGN#3