With　the　fast　development　of　E-commerce,　the　magnitudes　of　users　and　commodities　grow　dramastically,　resulting　in　the　extremely　sparse　user　rating　data.　Traditional　products＇　similarity　measurement　methods　perform　poorly　when　facing　sparse　user　rating　data.　Considering　the　extreme　sparsity　of　the　rating　data,　collaborative　filtering　algorithm　based　on　item　rating　prediction　is　introduced.　Meanwhile　collaborative　filtering　recommendation　technology　does　not　take　into　account　the　new　product　whose　rating　is　not　availble,　although　recommendation　value　is　high.　In　this　paper,　we　propose　an　improved　strategy,　which　uses　SVD　(Singular　Value　Decomposition)　matrix　decomposition　algorithm　and　cosine　similarity　to　group　users　into　clusters　with　common　interests　and　further　to　extract　the　eigenvector　of　the　commercial　products　to　be　evaluated　by　the　users　inside　each　group.　By　using　BP　(Back　Propagation)　neural　network　as　the　initial　training,　the　proposed　algorithm　can　predict　the　satisfaction　of　users　group　on　new　products.　For　those　new　products,　the　algorithm　assigns　higher　recommending　grade,　and　gives　the　priority　during　recommendation.　Finally.　the　results　of　this　optimized　collaborative　filtering　recommendation　algorithm　are　presented.　It　is　proven　that,　for　new　product　recommendation,　the　performance　of　the　new　algorithm　is　12%　better　than　that　of　the　traditional　collaborative　filtering　recommendation　algorithm.