LLMs hunt for surprising data patterns

🔥 What happened Lijia Liu and Hisashi Kashima (Kyoto University) propose DiSARM, a hybrid framework that scores association rules in transactional data for subjective interestingness using LLM world knowledge — while keeping statistical reliability and rule diversity intact. Tested on three datasets. 💡 Why it matters Pure LLM scoring doesn't scale and isn't statistically trustworthy. DiSARM pairs classic ARM mining with LLM judgment — a pragmatic middle path for EDA that makes surprise and diversity measurable instead of just counting frequent itemsets. ⚡ Our take Finally someone uses LLMs as a filter, not an oracle. Whether this beats a well-tuned lift metric in practice is the real question.