Traditional　recommendation　techniques　present　various　methods　to　measure　similarity　of　users　and　items　to　characterize　the　preferences.　However,　different　similarity　measure　focus　on　different　aspects　of　user-item　rating　list　and,　this　may　cause　incomplete　information　leveraged　by　similarity　measure　in　users＇　preference　characterization　leading　to　low　accuracy　on　recommendation.　This　paper　proposes　a　deep　learning　approach,　i.e.　DeRec,　to　learn　the　latent　item　association　from　user-item　rating　list　directly　for　predictive　recommendation　without　employing　a　similarity　measure.　The　loss　of　each　item　is　weighted　by　its　historical　probability　rated　by　users＇　past　preferences,　in　which　a　deep　learning　neural　network　is　adopted　to　predict　a　user＇s　potential　interest　on　the　items　using　the　user＇s　historical　items　as　input.　We　also　develop　two　strategies　to　produce　input　vectors　and　output　vectors　as　sampling　by　random　(Ran-Strategy)　and　sampling　by　distribution　(Pro-Strategy)　to　train　the　deep　neural　network　with　considering　the　sequential　characteristics　of　items　rated　by　users.　Experiments　on　the　App　dataset　and　the　MovieLens　dataset　demonstrate　that　the　proposed　DeRec　approach　outperforms　traditional　collaborative　filtering　methods　in　recommending　Apps　and　movies　in　both　MAP　and　MRR　measures.