Link prediction on graphs with different node types

Is it possible to use DGL to perform link prediction on a graph with multiple node types. And to restrict the LP network to predict links within only one of those types?

I am trying to model an online user interaction graph, and my target is to predict if user X is the same as user Y given their identifiers, page views, and other interactions. Those interactions will be represented as different node types on the graph to capture common user behavior.

I have read about the lack of support of heterogeneous graphs however not sure if that applies to my case. Any guidance will be highly appreciated.


1 Like

Hello Baddar,

It seems your model requires heterogeneous graph API and we are working on this.
Here is the heterogeneous graph API we proposed:

We are also working on an implementation of this API particularly for bipartite graphs. We should be able to have a preview for bipartite graphs very soon. I think your model might be implemented with bipartite graph APIs. We’ll keep you updated with the progress of our development. It’ll be great if you can provide us feedbacks during or after our development.



I wonder if at the current implementation you can restrict negative sampling to include a specific node type.

I also believe bipartite will not necessarily fit my use case as the different node types interact freely with each other.

Hi, it would be very nice if APIs for bipartite graphs can be used soon.
It seems that the heterograph branch in github is being developed.
So how many days can we probably use the preview version of bipartite graphs since we have an urgent project which will require GCN model for bipartite graphs.