As　both　rough　sets　theory　and　neural　network　in　data　mining　have　special　advantages　and　exiting　problems,　this　paper　presented　a　combined　algorithm　based　rough　sets　theory　and　BP　neural　network.　This　algorithm　deducts　data　from　data　warehouse　by　using　rough　sets＇　deduct　function,　and　then　moves　the　deducted　data　to　the　BP　neural　network　as　training　data.　By　data　deduct,　the　expression　of　training　will　become　clearer,　and　the　scale　of　neural　network　can　be　simplified.　At　the　same　time,　neural　network　can　easy　up　rough　set＇s　sensitivity　for　noise　data.　This　paper　presents　a　cost　function　to　express　the　relationship　between　the　amount　of　training　data　and　the　precision　of　neural　network,　and　to　supply　a　standard　for　the　change　from　rough　set　deduct　to　neural　network　training.