In　this　paper,　we　propose　an　adaptive　method　for　recommender　system　based　on　users＇　preference　to　items　represented　by　the　ratings　of　users.　This　method　defines　a　term-association　matrix　to　describe　the　relation　between　tags　and　items　properties.　A　gradient　descent　method　is　employed　to　compute　the　association　matrix.　The　association　matrix　is　then　used　to　implement　the　two　kinds　of　recommendation,　namely,　tag　recommendation　and　items　properties　recommendation.　©　2012　Springer-Verlag.