By letting your model learn the graph on which it operates, you may discover new insights about it. The way it re-arranges it can tell you more than the original alone.

A subway network recovered purely from passenger flow tells you something real about the city.

The adjacency matrix and its corresponding directed, weighted graph. Hover a cell, an edge, or a node to see the correspondence.

However, striking a balance between model growth and graph interpretability can be difficult. Some models can learn graphs that measurably improve their predictions while no longer being human-readable.