Recommender　systems　have　been　comprehensively　analyzed　in　the　past　decade　and　made　great　achievement　in　various　fields.　Generally　speaking,　the　recommendation　of　information　of　interests　is　based　on　the　potential　connections　among　users　and　items　implied　in　＇User-Item　Matrix＇.　However,　the　existing　algorithm　for　recommendation　will　be　degraded　and　ever　fail　in　the　case　of　sparseness　of　matrix.　To　resolve　this　problem,　a　new　algorithm　called　B-NMF　(blocks-coupled　non-negative　matrix　factorization)　is　proposed　in　this　paper.　With　this　algorithm:　(1)　the　reconstruction　performance　of　matrix　of　extreme　sparseness　is　improved　as　a　result　of　blocking　the　matrix　and　modeling　based　on　full　use　of　the　coupling　between　blocks;　(2)　the　coupling　between　different　blocks　is　ensured　via　a　coupling　mechanism　that　imposes　constraints　on　consistency　as　the　matrix　is　decomposed.　In　addition,　we　provide　an　approach　to　exploiting　homophily　effect　in　prediction　via　homophily　regularization　and　thus,　the　coupling　between　blocks　is　improved　via　extra　homophily　regularization　constraints.　Experiment　results　show　that　our　solution　is　superior　to　the　existing　ones　in　dealing　with　the　problem　of　extremely　sparse　matrix.　(C)　2017　Elsevier　B.V.　All　rights　reserved.