Attribute　graph　clustering　is　an　important　tool　to　analyze　and　understand　complex　networks.　In　recent　years,　graph　attention　auto-encoder　has　been　applied　to　attribute　graph　clustering　as　a　learning　method　for　unsupervised　feature　representation.　However,　the　graph　attention　auto-encoder　only　learns　the　feature　representation　of　the　nodes,　and　needs　to　use　traditional　clustering　algorithms　such　as　k-means　and　spectral　clustering　to　achieve　the　final　clustering　of　nodes.　During　the　optimization　process,　the　clustering　loss　cannot　be　fed　back　to　the　auto-encoder　and　the　extracted　features　are　not　necessarily　suitable　for　downstream　clustering　tasks　because　the　auto-encoder　model　for　feature　learning　and　the　clustering　model　are　mutually　independent.　To　overcome　this　problem,　we　propose　a　cluster-guided　graph　attention　auto-encoder　(CGATAE),　which　introduces　a　cluster-guided　pairwise　feature　relationship　preservation-based　non-negative　matrix　factorization　model　(FR-NMF)　into　the　graph　attention　auto-encoder.　The　model　CGATAE　obtains　the　final　clustering　results　while　learning　the　cluster-oriented　node　feature　representation.　Experiments　on　five　public　attribute　graph　datasets　verify　the　effectiveness　of　the　CGATAE　model,　and　its　clustering　quality　is　significantly　better　than　the　original　graph　attention　auto-encoder　model.