Consider removing low probability attribute values from nodes to speed up calculations (less attribute values to compute over).

[cmaclell]

The idea would be that when attribute values dip below a certain probability then they get removed. Not sure on any more specifics, but wanted to put the idea on the list.