[Objectives]　Urban　functional　regions　result　from　the　complex　interactions　and　mutual　influences　of　urban　planning　and　human　activities.　Accurately　identifying　these　regions　is　crucial　not　only　for　optimizing　the　allocation　of　public　resources,　such　as　infrastructure　and　services,　but　also　for　improving　the　efficiency　of　commercial　and　economic　activities　within　the　city.　In　recent　years,　the　rise　of　social　sensing　big　data　has　opened　up　new　possibilities　for　identifying　urban　functional　regions.　While　many　studies　have　leveraged　this　emerging　data　for　identification　purposes,　they　often　fall　short　in　fully　exploiting　the　deep　features　within　the　data.　Additionally,　they　frequently　fail　to　capture　and　utilize　the　complex　interrelationships　and　correlations　between　different　features,　leading　to　lower　identification　accuracy.　[Methods]　To　address　these　problems,　this　study　proposes　a　framework　for　identifying　urban　functional　regions　by　integrating　regional　embedding　representations　using　a　multi-head　attention　mechanism.　The　framework　leverages　mobile　phone　location　data　and　Point　of　Interest　(POI)　data,　employing　the　Node2vec　algorithm　to　extract　spatial　interaction　features　across　six　time　periods　(weekdays　and　weekends),　and　utilizing　the　GloVe　model　to　extract　semantic　features　of　regions.　Subsequently,　multi-head　attention　mechanisms　are　applied　to　effectively　integrate　these　features,　enabling　the　classification　and　identification　of　functional　areas　with　the　assistance　of　partially　labeled　data,　thereby　ensuring　greater　accuracy　and　reliability.　An　empirical　study　was　conducted　within　the　Third　Ring　Road　of　Fuzhou　City.　[Results]　The　results　demonstrate　that　the　proposed　method　generates　regional　representations　with　high　discriminative　power,　effectively　identifying　six　types　of　functional　areas.　The　Overall　Accuracy　(OA)　of　the　model　is　81%,　with　a　Kappa　coefficient　of　0.77.　[Conclusions]　Compared　to　the　DTW_KNN　and　Word2Vec　methods,　the　proposed　approach　improves　accuracy　by　30%　and　20%,　respectively,　through　fully　exploiting　the　global　spatial　interaction　and　semantic　features.　Furthermore,　ablation　experiments　confirm　the　superiority　of　the　proposed　method　in　integrating　multi-source　data　using　the　multi-head　attention　mechanism,　which　captures　complex　relationships　within　and　between　regions　while　assigning　higher　weights　to　critical　features.　This　results　in　an　improvement　of　approximately　18%　and　6%　in　OA　compared　to　single-source　data,　and　a　13%　improvement　compared　to　simple　data　fusion　methods.　Notably,　the　method　demonstrates　significant　advantages　in　identifying　residential　and　mixed-use　areas.　This　framework　provides　theoretical　support　for　the　application　of　multi-source　data　fusion　in　urban　research,　enriches　the　theoretical　system　of　urban　functional　regions　identification,　and　offers　important　support　for　promoting　the　intensive　use　of　urban　land　and　the　optimal　allocation　of　public　resources.　©　2025　Science　Press.　All　rights　reserved.