The　reduction　of　attributes　is　a　critical　problem　in　the　rough　set　theory.　Finding　the　minimal　reduct　is　turned　out　to　be　a　NP-hard　problem.　Many　heuristic　algorithms,　which　use　the　significance　of　the　condition　attribute　with　reference　to　the　decision　attributes　as　the　indication　for　attribute　selection,　have　been　proposed　in　this　area.　In　this　paper　the　pair-wise　complementarity　of　condition　attributes　is　defined　based　on　conditional　information　entropy　and　employed　as　a　heuristic　in　the　attribute　reduction　process.　Finally,　a　heuristic　algorithm　of　reduction　is　proposed　and　tested　on　the　UCI　machine　learning　repository.　It　can　be　verified　by　the　experimental　results　that　the　proposed　algorithm　is　feasible　and　effective.