Collaborative　Filtering(CF)　is　a　widely　used　technique　in　Recommender　System.　With　recent　development　in　deep　learning,　Neural　network　based　CF　has　gained　great　attention　in　recent　years,　especially　auto-encoders.　However,　the　main　disadvantage　of　autoencoder　based　CF　is　the　problem　of　the　large　sparse　target.　In　this　paper,　we　propose　a　training　strategy　to　tackle　this　issue,　We　run　experiments　on　two　popular　real　world　datasets　MovieLens　1M　and　MovieLens　10M.　Experiments　show　orders　of　magnitude　speed　up　while　Attaining　similar　accuracy　compare　to　existing　autoencoder　based　CF　method.　©　2016　The　authors　and　IOS　Press.　All　rights　reserved.