Data　envelopment　analysis　(DEA)　heavily　depends　on　the　dimensionality　of　the　variables,　and　previous　studies　address　the　problem　by　decreasing　the　dimensionality　with　a　minimal　loss　of　information.　Since　the　lost　information　can　also　have　the　impact　on　the　evaluation　performance,　this　paper　accordingly　proposes　an　approach　to　improve　the　discriminatory　power　of　DEA　without　losing　any　variables　information　and　without　requiring　any　additional　preferential　information.　Furthermore,　an　accelerating　approach　based　on　the　concept　of　parallel　computing　is　introduced　to　solve　the　multi-subsets　problem.　Listing　all　the　possible　variables　subsets　as　the　nodes,　then　the　DEA　efficiencies　under　each　node　are　calculated,　and　the　corresponding　purity　of　information　can　be　scientifically　generated　based　on　Renyi＇s　entropy.　Subsequently,　the　important　degrees　of　nodes　are　obtained　by　normalizing　the　purities　of　information,　and　the　comprehensive　efficiency　scores　can　be　finally　generated　by　the　weighted　sum　between　the　important　degrees　and　the　efficiencies　under　the　corresponding　node.　Two　specific　examples　are　provided　to　evaluate　the　performance.