The underlying problem is that the “real” graph of relations exists in a high-dimensional space; ML “embeddings” tend to tease out the actual structure of this

The underlying problem is that the “real” graph of relations exists in a high-dimensional space; ML “embeddings” tend to tease out the actual structure of this space. Once you have a good high-dimensional representation, you can use something like t-SNE to reduce that to 2 (or 3) dimensions in a way that makes sense. But as you noted, if you use all the nodes, this dimensionality reduction loses quite a bit of information. The solution is to have a way of interactively limiting the number of nodes in an exploratory way, and having a dynamically-applied t-SNE running on only that subset of nodes. If you set the iterative parameters right, you can get some really fun interactive physics-like visualizations where nodes attract and repulse each other in intuitive ways. (“This thing is like this other thing in one dimension, but also not like it in another dimension” kind of stuff.)