Column-stores gained popularity as a promising physical design alternative. The main overhead of query processing in Column-stores is on-the-fly tuple reconstruction for multi-attributes queries. Typical column-stores, such as C-Store and MonetDB, use projections to support tuple reconstruction. But how to select attributes for a projection is keeping open problem. This paper presents an adaptive approach to solve this problem. Our approach exploits an adaptive algorithm to cluster attributes for each projection. We show that our approach can conform well to clustering attributes for projections in column-stores, and it enable projections to adapt dynamically to users query customs, therefore is an effective method to select attributes for projections.