Show HN: Association rule mining on 21.6M poker hands

**Show HN: Association rule mining on 21.6M poker hands**

We built this as a side project that grew out of something completely different.I work on ET-Miner [https://zenodo.org/records/18674353], which is a GPU-accelerated frequent itemset mining pipeline based on the infamous apriori-algorithm. We came with the idea to reformulate the algorithm into a fully vectorized implementation, using a boolean transaction matrix representation,CUDA kernels + Rust group builder for index construction to speed up computations. The original use case was mining protein structure patterns from AlphaFold, where we processed 109.2M proteins and extracted 16.8 billion frequent itemsets for protein structural motif discovery. At some point I realized the same pipeline could be pointed at any domain with structured categorical data, so I pointed it at poker, one of my long-standing hobbies.What we learned: Most of the "surprising" patterns the mining surfaces are things good players already know intuitively: positional advantages, aggression frequency correlations, stack-to-pot ratios. But seeing them as statistically validated itemsets with exact support counts is different from folk wisdom...

---

**[devsupporter 해설]**

이 기사는 Show HN에서 제공하는 최신 개발 동향입니다. 관련 도구나 기술에 대해 더 알아보시려면 원본 링크를 참고하세요.