GRACE Improves Mixed-Data Clustering with LLM Semantics

Clustering mixed tabular data, which combines continuous numerical measurements with discrete categorical symbols, presents a significant challenge due to the inherent heterogeneity between these data types. Traditional clustering algorithms often rely solely on internal dataset statistics to infer relationships between categorical values, potentially missing conceptually obvious but statistically unobserved affinities. While large language models (LLMs) offer vast external world knowledge, applying their text-centric reasoning to abstract tabular concepts has been difficult. Existing methods often embed LLMs into iterative metric learning loops, leading to computationally intensive processes that compromise scalability for semantic enrichment. To overcome this, researchers propose GRACE (LLM-Grounded Semantic Metric Spaces for Scalable Mixed-Data Clustering). GRACE innovatively shifts the semantic acquisition to the attribute-value level, using a multi-perspective LLM querying strategy to map heterogeneous values into knowledge-informed descriptions. This "one-shot grounding" creates general-purpose semantic representations that embed diverse attributes into a unified space, crucially decoupling the expensive LLM invocation from the iterative optimization process. GRACE also cross-validates these external semantics with internal dataset statistics, ensuring alignment with the specific cluster structure. The result is a framework that matches the scalability of conventional methods while achieving superior clustering accuracy and conceptual interpretability across various benchmarks.