Federated　Graph　Learning　(FGL)　is　a　fusion　of　Federated　Learning　(FL)　and　Graph　Neural　Network.　It　is　well　known　that　one　of　the　major　challenges　that　FL　needs　to　address　is　the　non-IID　problem.　However,　FGL　not　only　needs　to　address　the　non-IID　problem　at　the　data　feature　level　but　also　at　the　structure　feature　level.　In　addition,　addressing　unbalanced　engagement　among　different　clients　in　global　training　is　crucial　for　further　advancing　the　FGL　application.　Aiming　to　address　the　aforementioned　challenges,　we　propose　a　Soft　Clustering　Based　Federated　Graph　Learning　(SCFGL)　on　non-IID　Graphs,　which　mitigates　the　non-IID　problem　of　both　data　and　structure　feature　levels　while　balancing　the　impact　of　clients　with　different　levels　of　engagement.　Firstly,　to　address　the　non-IID　distribution　of　structural　features　among　graph-structured　data　from　different　clients,　we　develop　a　Truncated　Singular　Value　Decomposition　(TSVD)　based　soft　clustering　method,　which　constructs　principal　vectors　by　applying　TSVD　to　the　structural　embeddings　and　divides　personalized　clusters　by　structural　similarity.　Secondly,　we　design　a　model　aggregation　method　that　takes　into　account　both　the　level　of　client　engagement　and　the　model　quality　to　tackle　unbalanced　client　engagement　in　real-world　application　scenarios.　Finally,　we　conduct　extensive　experiments　under　various　settings,　such　as　cross-dataset　and　cross-domain.　The　results　demonstrate　that　our　proposed　method　outperforms　existing　FGL　state-of-the-art　methods　in　terms　of　performance.　©　2025　Elsevier　Ltd