Mobile　Crowdsensing　(MCS)　is　a　sensing　paradigm　that　enables　large-scale　smart　city　applications,　such　as　environmental　sensing　and　traffic　monitoring.　However,　traditional　MCS　often　suffers　from　performance　degradation　due　to　the　limited　spatiotemporal　coverage　of　collected　data.　In　this　context,　Sparse　MCS　has　been　proposed,　which　utilizes　data　inference　algorithms　to　recover　full　data　from　sparse　data　collected　by　users.　However,　existing　Sparse　MCS　approaches　often　overlook　spatiotemporal　fractures,　where　no　data　is　observed　either　for　a　sensing　subarea　across　all　sensing　time　slots　(temporal　fracture),　or　for　a　sensing　time　slot　in　all　sensing　subarea　(spatial　fracture).　Such　spatiotemporal　fractures　pose　great　challenges　to　the　data　inference　algorithms,　as　it　is　difficult　to　capture　the　complex　spatiotemporal　correlations　of　the　sensing　data　from　very　limited　observations.　To　address　this　issue,　we　propose　a　Graph-and　Attention-based　Matrix　Completion　(GAMC)　method　for　the　spatiotemporal　fracture　data　inference　problem　in　Sparse　MCS.　Specifically,　we　first　pre-fill　the　general　missing　values　using　the　classical　Matrix　Factorization　(MF)　technique.　Then,　we　propose　a　neural　network　architecture　based　on　Graph　Attention　Networks　(GAT)　and　Transformer　to　capture　complex　spatiotemporal　dependencies　in　the　sensing　data.　Finally,　we　recover　the　complete　data　with　a　projection　layer.　We　conduct　extensive　experiments　on　three　real-world　urban　sensing　datasets.　The　experimental　results　show　the　effectiveness　of　the　proposed　method.　IEEE