Emotion　recognition　is　widely　used　in　many　areas,　such　as　medicine　and　education.　Due　to　the　obvious　difference　in　duration　and　intensity　between　micro　and　macro　expression,　the　same　model　cannot　be　used　to　classify　emotions　precisely.　In　this　paper,　an　algorithm　for　emotion　recognition　based　on　graph　neural　network　is　proposed.　The　proposed　method　involves　four　key　steps.　Firstly,　data　augmentation　is　used　to　increase　the　diversity　of　original　data.　Secondly,　graph　network　is　built　based　on　feature　points.　The　feature　points　Euclidean　distance　is　calculated　as　the　initial　value　of　the　matrix.　Thirdly,　Laplacian　matrix　is　obtained　according　to　the　matrix.　Finally,　graph　neutral　network　is　utilized　to　bridge　the　relationship　between　feature　vectors　and　emotions.　In　addition,　a　new　dataset　named　FEC-13　is　provided　by　subdivided　traditional　six　kinds　of　emotions　to　thirteen　categories　according　to　the　intensity　of　emotions.　The　experimental　results　show　that　a　high　accuracy　is　reached　with　a　small　amount　of　training　data,　especially　CASME　II　dataset,　which　achieves　an　accuracy　of　95.49%.　A　cross-database　study　indicates　that　proposed　method　has　high　generalization　performance　and　the　accuracy　of　FEC-13　dataset　is　74.99%.　©　2021,　Springer　Nature　Singapore　Pte　Ltd.