dgraph stores predicates per group and storage does not cross groups, so could you not have your europe-only predicate on a group that does not have singapore learner nodes attached, and have the globally used predicates on groups that do have learner nodes? Obviously this would be manual instead of automagic based on usage, but seems like that meets your use case.
The automatic rebalancing system in dgraph is based on disk usage per predicate, and I feel like a better algorithm could be used to distribute based on usage instead of disk size, but that may be a different issue all together.
A different tangential feature but may be interesting here is virtual tablets, like those implemented by cassandra/scylla (and slightly different style but cockroachdb) to split up the storage of a tablet into virtual segments and replicate those as opposed to keeping all values for tablet X in one group, no exceptions.