If you are processing PS-like targets, use the star graph.
If you are dealing with distributed targets, it’s better to use a multi-master graph.
The MST is theoretically interesting, but practically it’s better to use more connections to stabilize the inversion.
If you use the (spatial) coherence as weight, it’s strongly suggested to use a multi-master graph.
Also, when you have few images, it’s better to use many redundant interferograms.