Learning Better Representations From Less Data For Propositional Satisfiability