You can use the --esampmin option with the Dgraph process, to specify the minimum number of records to sample during refinement computation.
This option is useful because sampling the entire navigation state during the refinement computation can be one of the more performance intensive operations for the Dgraph.
For most applications, larger values for --esampmin reduce performance without improving the quality of refinement ordering. For some applications with extremely large, non-hierarchical attributes (if they cannot be avoided), larger values can meaningfully improve refinement ordering quality with minor performance cost.