Stable Decision Trees: Using Local Anarchy for Efficient Incremental Learning