In　recent　years,　the　use　of　convolutional　neural　network　(DNN)　for　depression　recognition　has　received　a　lot　of　research.　However,　DNN　can　only　be　employed　for　the　modelling　of　video,　audio　and　natural　language　processing,　and　is　not　suitable　for　learning　with　few　samples　and　tabular　data.　In　this　paper,　for　tabular　data　based　few　shot　learning,　we　propose　a　multiple　parallel　graph　attention　networks　(pGAT)　architecture.　As　the　first,　calculate　information　of　multiple　emotional　bandwidths　(such　as　information　entropy,　energy)　based　on　the　pupil　size,　and　extract　classification　features　according　to　their　statistical　distribution,　and　then,　distance　similarity　(Euclid,　Manhattan,　Chebyshev)　is　used　to　construct　three　pGAT,　finally,　fuse　the　three　streams　for　classifying　depression.　The　results　show　that　the　classification　sensitivity　and　specificity　are　84.88%　and　83.16%,　respectively,　which　have　better　recognition　performance　than　the　related　research　recently.