I am aware of papers where the graph is assumed to be known (e.g. arxiv.org/abs/2103.06392), and indeed I can construct one from the data that I have. But my concern is whether using the empirical graph ignores selection bias from customers choosing to form edges with certain products.

the graph is assumed to be fixed in this line of work, right? so you wouldn't worry about selection bias re: edge formation per se, since counterfactuals are over treatments and not the structure of the graph. basing this on JP's slides docs.google.com/presentation...