A　community　is　composed　of　closely　related　nodes.　Detecting　communities　in　a　network　has　many　practical　applications,　such　as　online　product　recommendation,　biological　molecule　discovery　and　criminal　group　tracking.　In　recent　years,　network　representation　learning　(NRL)　has　attracted　much　attention　in　the　field　of　community　detection　because　it　can　effectively　extract　complex　relations　between　nodes　which　improves　the　quality　of　detected　communities.　In　many　real-world　networks,　the　rich　attribute　information　contained　in　nodes　and　the　similarity　between　nodes　and　their　multi-order　neighbors　has　significant　contributions　to　the　generation　of　node　embedding　vectors　in　NRL.　However,　existing　NRL　algorithms　treat　a　node＇s　different　order　neighbors　equally　as　high-order　contexts,　leading　to　the　ignorance　of　their　different　impacts　on　the　generation　of　the　node＇s　embedding　vector.　In　addition,　these　algorithms　do　not　focus　on　the　coupling　and　interaction　relations　between　nodes　playing　similar　structural　roles,　which　may　ignore　some　nodes　in　a　community　with　similar　structural　roles.　In　this　paper,　we　propose　a　novel　autoencoder　considering　the　multi-order　similarity　and　structural　role　similarity　(AMOSOS)　to　solve　the　above　problems.　First,　we　design　a　strategy　to　obtain　a　multi-order　weight　matrix　which　preserves　the　differential　influence　of　neighbors　of　different　orders　by　sequentially　decreasing　the　weight　of　each　order.　Second,　we　design　a　role　similarity　indicator　to　capture　the　complex　coupling　and　interaction　relations　of　nodes　in　the　network.　Experimental　results　on　synthetic　networks　and　real-world　networks　show　that　our　proposed　algorithm　is　more　accurate　than　existing　network　representation　learning　algorithms　for　the　task　of　community　detection.