Graph　representation　learning　is　a　crucial　area　in　machine　learning,　with　widespread　applications　in　social　networks,　recommendation　systems,　and　traffic　flow　prediction.　Recently,　Graph　Transformers　have　emerged　as　powerful　tools　for　this　purpose,　garnering　significant　attention.　In　this　work,　we　observe　a　fundamental　issue　of　previous　Graph　Transformers　that　they　overlook　the　scale-related　information　gap　and　often　employ　an　identical　attention　computation　method　for　different-scale　node　interactions,　leading　to　suboptimality　of　model　performance.　To　address　this,　we　propose　a　Multi-Scale　Attention　Graph　Transformer　(MSA-GT)　that　enables　each　node　to　conduct　adaptive　interactions　conditioned　on　different　scales　from　both　local　and　global　perspectives.　Specifically,　MSA-GT　guides　several　attention　mechanisms　to　focus　on　individual　scales　and　then　perform　customized　combinations　via　an　attention-based　fusion　module,　thereby　obtaining　much　more　semantically　fine-grained　node　representations.　Despite　the　potential　of　the　above　design,　we　still　observe　over-　fitting　to　some　extent,　which　is　atypical　challenge　for　training　Graph　Transformers.　We　propose　two　additional　technical　components　to　prevent　over-fitting　and　improve　the　performance　further.　We　first　introduce　a　path-　based　pruning　strategy　to　reduce　ineffective　attention　interactions,　facilitating　more　accurate　relevant　node　selection.　Additionally,　we　propose　a　Heterophilous　Curriculum　Augmentation　(HCA)　module,　which　gradually　increases　the　training　difficulty,　forming　a　weak-to-strong　regularization　schema　and　therefore　enhancing　the　model＇s　generalization　ability　step-by-step.　Extensive　experiments　show　that　our　method　outperforms　many　state-of-the-art　methods　on　eight　public　graph　benchmarks,　proving　its　effectiveness.