Collaborative　filtering　(CF)　is　a　widely　used　technique　in　recommender　systems.　With　rapid　development　in　deep　learning,　neural　network-based　CF　models　have　gained　great　attention　in　the　recent　years,　especially　autoencoder-based　CF　model.　Although　autoencoder-based　CF　model　is　faster　compared　with　some　existing　neural　network-based　models　(eg,　Deep　Restricted　Boltzmann　Machine-based　CF),　it　is　still　impractical　to　handle　extremely　large-scale　data.　In　this　paper,　we　practically　verify　that　most　non-zero　entries　of　the　input　matrix　are　concentrated　in　a　few　rows.　Considering　this　sparse　characteristic,　we　propose　a　new　method　for　training　autoencoder-based　CF.　We　run　experiments　on　two　popular　datasets　MovieLens　1 M　and　MovieLens　10 M.　Experimental　results　show　that　our　algorithm　leads　to　orders　of　magnitude　speed-up　for　training　(stacked)　autoencoder-based　CF　model　while　achieving　comparable　performance　compared　with　existing　state-of-the-art　models.　©　2018　John　Wiley　&　Sons,　Ltd.