To　solve　the　problem　of　low　recommendation　accuracy　caused　by　too　little　user　behavior　information　in　the　current　behavior　recommendation　system,　an　algorithm　based　on　end-to-end　data　enhancement　was　proposed.　In　this　paper,　knowledge　graph　is　constructed　by　learning　and　integrating　structured　knowledge　network.　Moreover,　the　characteristics　of　users　with　high　preference　similarity　can　be　propagated　through　the　inter-entity　relations　mapped　by　the　knowledge　map　to　reconstruct　the　preference　vector　of　users.　Through　comparative　experiments　on　open　data　sets,　the　AUC　of　RNN　model,　CNN　model,　RNN　attention　model　and　ATRank　were　improved　by　3.28%,　2.35%,　2.89%　and　1.30%,　respectively.　©　2019　IEEE.