With　increasing　popularity　and　larger　real-world　applicability,　graph　self-supervised　learning　(GSSL)　can　significantly　reduce　labeling　costs　by　extracting　implicit　input　supervision.　As　a　promising　example,　graph　masked　auto-encoders　(GMAE)　can　encode　rich　node　knowledge　by　recovering　the　masked　input　components,　e.g.,　features　or　edges.　Despite　their　competitiveness,　existing　GMAEs　focus　only　on　neighboring　information　reconstruction,　which　totally　ignores　distant　multi-hop　semantics　and　thus　fails　to　capture　global　knowledge.　Furthermore,　many　GMAEs　cannot　scale　on　large-scale　graphs　since　they　suffer　from　memory　bottlenecks　with　unavoidable　full-batch　training.　To　address　these　challenges　and　facilitate　“high-level”　discriminative　semantics,　we　propose　a　simple　yet　effective　framework　(i.e.,　HopMAE)　to　encourage　hop-perspective　semantic　interactions　by　adopting　multi-hop　input-rich　reconstruction　while　supporting　mini-batch　training.　Despite　the　rationales　of　the　above　designs,　we　still　observe　some　limitations　(e.g.,　sub-optimal　generalizability　and　training　instability),　potentially　due　to　the　implicit　gap　between　the　task-triviality　and　input-richness　of　reconstruction.　Therefore,　to　alleviate　task-triviality　and　fully　unleash　the　potential　of　our　framework,　we　further　propose　a　combined　fine-grained　loss　function,　which　generalizes　the　existing　ones　and　significantly　improves　the　difficulties　of　reconstruction　tasks,　thus　naturally　alleviating　over-fitting.　Extensive　experiments　on　eight　benchmarks　demonstrate　that　our　method　comprehensively　outperforms　many　state-of-the-art　counterparts.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.