The　random-walk-based　attribute　network　embedding　methods　aim　to　learn　a　low-dimensional　embedding　vector　for　each　node　considering　the　network　structure　and　node　attributes,　facilitating　various　downstream　inference　tasks.　However,　most　existing　attribute　network　embedding　methods　base　on　random　walk　usually　sample　many　redundant　samples　and　suffer　from　inconsistency　between　node　structure　and　attributes.　In　this　paper,　we　propose　a　novel　attributed　network　embedding　method　for　community　detection,　which　can　generate　node　sequences　based　on　attributed-subgraph-based　random　walk　and　filter　redundant　samples　before　model　training.　In　addition,　an　improved　network　embedding　enhancement　strategy　is　applied　to　integrate　high-order　attributed　and　structure　information　of　nodes　into　embedding　vectors.　Experimental　results　of　community　detection　on　synthetic　network　and　real-world　network　show　that　our　algorithm　is　effective　and　efficient　compared　with　other　algorithms.