It’d be good to see someone using this to make a node or edge prediction.
I can kind of follow this but would need to copy/paste it into my app and play around to fully get it, but can see it follows the approach used in the Python-based video you posted above.
It’d possibly be simplest if the person who wrote GCNs for tensorflow could just port it to tensorflow.js
In my scenario I have 61 node types, and 11 edge types, and both nodes + edges can have properties which are important in making node and edge predictions, e.g.
- node prediction - given a source node + edge, what is the target node predicted?
- edge prediction - given a source and target node, what is the edge predicted?
My training set is a bunch of graphs of varying sizes, and containing different subsets of the 61/11 nodes/edges respectively, and for training I need to train on multiple graphs, i.e. there isn’t just 1 big graph.
So in my “Simple GCN” implementation I decompose the, say, 20 graphs, all into a bunch of tripes (source, edge, target) and put it through a sequential model.
This seems to allow for pretty good node/edge predictions.
I can’t see how in the video or link you provided you would go about training on multiple graphs – I could treat them as 1 big graph where the ‘subgraphs’ weren’t connected to each other I guess.