With　the　advent　of　the　era　of　big　data,　the　traditional　recommendation　system　mainly　for　a　single　user,　but　with　the　society　and　the　rapid　development　of　e-commerce,　more　and　more　people　participate　in　activities　together,　in　the　form　of　multiple　users　and　groups　of　recommendation　system　recommended　service　object　by　a　single　user　extensions　for　group　members,　has　become　a　hot　topic　in　the　current　society.　In　view　of　the　low　accuracy　rate　of　group　recommendation　and　the　inconsistent　fusion　strategy　among　group　members,　the　traditional　solution　method　is　matrix　decomposition.　MF　USES　a　simple　and　fixed　inner　product　to　estimate　the　complex　user-project　interaction　in　low-dimensional　potential　space,　which　will　cause　the　problem　of　limitation.　Therefore,　a　group　recommendation　algorithm　combining　self-attention　and　NCF　to　solve　the　problem　of　group　preference　fusion　is　proposed.　We　use　neural　network　to　learn　the　group　recommendation　of　fusion　strategy　through　self-attention,　and　further　integrate　the　user-project　interaction　improvement　group　recommendation　through　NCF　model.　The　self-attention　mechanism　provided　in　this　paper　was　verified　on　qunar　and　CAMRA2011　data　sets.　Compared　with　other　common　fusion　strategies,　the　overall　average　performance　of　the　proposed　mechanism　in　NDCG　and　HR　was　improved.　©　2020　IEEE.