Semantic-Aware Association Rule Mining at Scale: Ontology-Free Generalization with LLMs

Classical association rule mining (ARM) discovers symbolic co-occurrence patterns from transactional datasets, but suffers from semantic brittleness: rules are tied to exact item identities, yielding overly specific and fragmented results. Existing generalization approaches rely on manually curated taxonomies, limiting scalability in dynamic, heterogeneous big data environments. We present Semantic-Aware Association Rule Mining (SA-ARM), a framework that integrates large language models (LLMs) into the ARM pipeline to enable ontologyfree semantic generalization. SA-ARM performs (i) LLMbased item normalization to denoise heterogeneous item labels, (ii) ontology-free semantic grouping of items into latent categories, and (iii) support normalization across abstraction levels to yield statistically meaningful, generalizable rules. This design overcomes the brittleness of exact-match ARM while adapting dynamically to evolving datasets. We evaluate SA-ARM on real-world transactional datasets, comparing item-level, embedding-based, and LLM-based groupings. Results show that LLM-driven grouping produces more compact, interpretable, and higher-coverage rules, while support normalization ensures statistical robustness. Beyond retail transactions, our framework applies broadly to domains where symbolic cooccurrence is mined at scale, including recommendation systems, biomedical data, and scientific knowledge discovery.