-
|
We are planning to explore PyG’s HeteroConv/HGTConv for our use case (similar to the example given in the original HGT paper, a dynamic heterogeneous graph). By analogy, we want to find PhDs authors at 2023, e.g., I am a bit confused how the HeteroConv/HGTConv example in the PyG maps to the example in the HGT paper in terms of “dynamic”. In the HGT paper it claims it can handle dynamic graph, e.g., associating an event node with different timestamps such as “WWW@2015”, “WWW@2016”, etc. Does it mean creating individual event node such as “WWW@2015”, “WWW@2016”, ... in the graph (as I don’t see such thing in the PyG example) ? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
Temporal information is added via relative positional encoding to the node representations, as discussed in Section 3.5. One can achieve this via the usage of |
Beta Was this translation helpful? Give feedback.
-
|
Since we need this feature in our current project, I could contribute an implementation of relative temporal encoding for HGTConv. Would this be beneficial to the project? |
Beta Was this translation helpful? Give feedback.
Temporal information is added via relative positional encoding to the node representations, as discussed in Section 3.5. One can achieve this via the usage of
torch_geometric.nn.PositionalEncoding. However, it is not (yet) integrated into ourHGTConvimplementation.