Graph　self-supervised　learning　(GSSL)　has　garnered　significant　attention　owing　to　its　powerful　expressive　capabilities.　It　learns　from　unlabeled　graph　data　through　pretext　tasks,　such　as　AttributeMask,　S2GRL,　and　GZL.　However,　most　existing　methods　face　challenges　when　performing　these　pretext　tasks:　(1)　The　model　tends　to　select　nodes　whose　features　are　easy　to　predict,　resulting　in　the　learned　information　being　too　simple　and　lacking　actual　semantic　value.　(2)　The　noise　in　node　features　can　mislead　the　model　to　learn　wrong　information,　which　can　easily　lead　to　over-fitting　and　reduce　generalization　ability.　To　address　these　limitations,　we　propose　a　novel　GSSL　framework　termed　Selection-Enhanced　GSSL　(SE-GSSL),　which　integrates　sample　selection　into　the　learning　paradigm.　Furthermore,　we　introduce　an　end-to-end　Soft-Mask-based　SE-GSSL　framework　as　a　concrete　implementation.　First,　we　design　multiple　individual　encoders　to　extract　differentiated　knowledge　of　node　features　and　graph　topology,　and　achieve　complementary　enhancement　in　fusion.　Second,　a　multi-head　global　projection　matrix　with　diversity　regularization　is　presented　to　calculate　the　scores　for　all　nodes　to　enhance　the　evaluation　of　node　importance.　Third,　a　new　masking　strategy　called　Soft-Mask　is　proposed　to　mask　nodes,　which　can　adaptively　adjust　the　difficulties　of　different　self-supervised　reconstruction　tasks.　Finally,　extensive　experiments　on　various　public　graph　benchmarks　demonstrate　that　the　proposed　framework　can　achieve　better　or　at　least　competitive　performance　compared　to　many　other　baselines　or　recent　state-of-the-art　methods　in　semi-supervised　node　classification　and　node　clustering　tasks.　©　2025　Elsevier　B.V.