Existing　methods　based　on　graph　convolutional　neural　network　have　made　some　achievements　in　graph　clustering,　but　most　of　them　do　not　consider　how　to　obtain　the　most　beneficial　embedding　representation　for　clustering　when　the　clustering　structure　cannot　be　predicted,　and　cannot　maximize　the　use　of　the　graph　structural　and　attributed　information　and　their　relationship.　In　this　paper,　we　propose　a　marginalized　graph　autoencoder　with　subspace　structure　preserving,　which　adds　a　self-expressive　layer　to　reveal　the　clustering　structure　of　node　attributes　based　on　the　marginalized　graph　autoencoder,　so　that　the　output　of　the　autoencoder　maintains　the　multi-subspace　structure　of　the　input　feature　matrix　and　matches　the　clustering　target,　resulting　in　a　cluster-oriented　graph　representation　to　improve　clustering　performance.　Experiments　on　four　public　datasets　show　that　the　algorithm　can　effectively　improve　the　performance　of　the　graph　clustering.　©　2023　IEEE.