UniSAGE framework unifies static & dynamic attributes, outperforms by 10%+

Traditional approaches to modeling data with both static attributes (like user demographics) and dynamic records (like purchase history) often require extensive manual design and treat these data types separately, missing important interactions. UniSAGE, proposed by researchers Taoran Fang and colleagues, solves this by introducing a unified graph-based framework. It first constructs a global attribute graph that captures hierarchical and temporal relationships in a single structure. To maintain representational consistency, it introduces two orthogonal parameter subspaces—one for static aggregation, one for dynamic reasoning—that operate within a shared semantic space. A lightweight hyper-structure mechanism then enables task-specific interactions between these subspaces, allowing the model to capture complex cross-attribute dependencies.

UniSAGE is fully automated and adapts to evolving data schemas, making it practical for real-world applications. The authors validated it on multiple public benchmarks and a real-world financial behavior dataset, where it consistently outperformed existing methods, often by over 10%. This work is especially relevant for domains like finance, e-commerce, and healthcare, where hierarchical and time-varying data are common. By unifying static and dynamic attributes without manual schema design, UniSAGE offers a scalable, generalizable solution for modeling complex heterogeneous data.