BigML uses the same calculations when pruning as for reporting a node's confidence and error. If a split doesn't reduce the parent node's estimated error, then BigML prunes it away.
There are three different pruning strategies you can define while configuring your model: Smart (called "selective_pruning" in the BigML API), Statistical and No statistical pruning. Please take a look at the Dashboard documentation to see which one suits your model best. If you are using the BigML API, read this section of the documentation for developers.