Group　activity　recognition,　a　challenging　task　that　requires　not　only　recognizing　the　individual　actions　of　each　person　but　also　inferring　relationships　among　persons,　has　received　considerable　attention.　Previous　methods　infer　coarse-level　relations　based　on　holistic　features　of　individuals.　This　paper　goes　one　step　further　beyond　existing　methods　by　learning　part-aware　spatial-temporal　graph　convolutional　network　(PSTGCN)　to　model　fine-grained　relations　using　part-level　features.　The　PSTGCN　includes　two　core　graphs,　the　locally-connected　graph,　and　the　fully-connected　graph.　The　locally-connected　graph　in　which　all　parts　of　the　same　person　are　connected　mines　structural　information　of　individuals.　The　fully-connected　graph　in　which　all　parts　of　persons　are　connected　extracts　the　latent　relationships　among　individuals　in　part-level.　Extensive　experiments　on　two　widely-used　datasets:　the　Volleyball　dataset　and　the　Collective　Activity　dataset,　are　conducted　to　evaluate　our　PSTGCN,　and　the　improved　performance　demonstrates　the　effectiveness　of　our　approach.　Visualizations　of　the　part-based　relation　graph　and　the　group-level　features　indicate　our　method　can　capture　the　discriminative　information　for　group　activity　recognition.　©　2021,　Springer　Nature　Switzerland　AG.