Music　content　has　recently　been　identified　as　useful　information　to　promote　the　performance　of　music　recommendations.　Existing　studies　usually　feed　low-level　audio　features,　such　as　the　Mel-frequency　cepstral　coefficients,　into　deep　learning　models　for　music　recommendations.　However,　such　features　cannot　well　characterize　music　audios,　which　often　contain　multiple　sound　sources.　In　this　paper,　we　propose　to　model　and　fuse　chord,　melody,　and　rhythm　features　to　meaningfully　characterize　the　music　so　as　to　improve　the　music　recommendation.　Specially,　we　use　two　userbased　attention　mechanisms　to　differentiate　the　importance　of　different　parts　of　audio　features　and　chord　features.　In　addition,　a　Long　Short-Term　Memory　layer　is　used　to　capture　the　sequence　characteristics.　Those　features　are　fused　by　a　multilayer　perceptron　and　then　used　to　make　recommendations.　We　conducted　experiments　with　a　subset　of　the　last.fm-1b　dataset.　The　experimental　results　show　that　our　proposal　outperforms　the　best　baseline　by　3.52%　on　HR@10.