Abstract

UniSAGE unifies hierarchical static attributes and dynamic records in a global attribute graph, combining orthogonal parameter subspaces with a lightweight hyper-structure for joint aggregation and reasoning. Across public benchmarks and a real-world financial behavior dataset, it consistently outperforms existing methods, with gains above 10% on several tasks.

Abstract

UniSAGE unifies hierarchical static attributes and dynamic records in a global attribute graph, combining orthogonal parameter subspaces with a lightweight hyper-structure for joint aggregation and reasoning. Across public benchmarks and a real-world financial behavior dataset, it consistently outperforms existing methods, with gains above 10% on several tasks.