The　purpose　of　community　detection　is　to　discover　closely　connected　groups　of　entities　in　complex　networks　such　as　interest　groups,　proteins　and　vehicles　in　social,　biological　and　transportation　networks.　Recently,　autoencoders　have　become　a　popular　technique　to　extract　nonlinear　relationships　between　nodes　by　learning　their　representation　vectors　through　an　encoder-decoder　neural　structure,　which　is　beneficial　to　discovering　communities　with　vague　boundaries.　However,　most　of　the　existing　autoencoders　take　restoring　a　network＇s　adjacency　matrix　as　their　objective,　which　puts　emphasis　on　the　first-order　relationships　between　the　nodes　and　neglects　their　higher-order　relationships　that　may　be　more　useful　for　community　detection.　In　this　paper,　we　propose　a　novel　attentional-walk-based　autoencoder　(AWBA)　which　integrates　random　walk　considering　attentional　coefficients　between　each　pair　of　nodes　into　the　encoder　to　mine　their　high-order　relationships.　First,　the　attention　layers　are　added　to　the　encoder　to　learn　the　influence　of　a　node＇s　different　neighbors　on　it　in　encoding.　Second,　we　develop　a　new　random　walk　strategy　that　embeds　the　attention　coefficients　and　the　community　membership　of　the　nodes　obtained　by　a　seed-expansion-based　clustering　algorithm　into　the　computation　of　the　transition　probability　matrix　to　instill　both　low　and　high　order　relationships　between　the　nodes　into　the　representation　vectors.　The　experimental　results　on　synthetic　and　real-world　networks　verify　the　superiority　of　our　algorithm　over　the　baseline　algorithms.