Finding the Weirdest Grocery Combos with 3M Instacart Orders and a Humor Index

What's the weirdest combination of items someone could buy at a grocery store? That's the question Roger Dickey set out to answer by digging into Instacart's open-sourced dataset of 3 million orders. The result is a fun technical exercise in data analysis, classification, and a dash of humor scoring.

From raw data to meaningful pairs

The dataset contains over 3.2 million orders with about 10 items per order, covering nearly 50,000 unique products across 134 aisles. The naive approach—just counting co-occurrences—fails because 97% of product pairs never appear together, and 22 million pairs appear exactly once. Sorting by raw frequency gives you boring results like bananas and mixed nuts.

Dickey switched to ranking by lift: actual co-occurrence divided by expected co-occurrence. This surfaced pairs that are statistically unlikely, but still not funny—like organic and non-organic bananas in the same cart.

Granularity is the enemy

The 50,000 products are too granular (28 condom SKUs, 110 apple variants), while 134 aisles are too broad. The fix was the GS1 Global Product Classification (GPC) ontology, which groups products into ~1,700 bricks (e.g., chanterelles, morels). A two-stage classifier—embeddings via qwen3-embedding:8b for shortlisting, then gpt-4.1-mini for final selection—mapped all products to bricks at a cost of ~$5.

With bricks, the lift-based analysis finally produced non-obvious pairs: wine and frozen grains, spirits and dates, beer and baby leaves.

Adding a humor index

But statistically rare doesn't mean funny. Dickey had Claude score each of the ~1,000 grocery bricks on a 0–1 humor scale based on taboo or comedic potential. Condoms, intimate lubricants, and enemas scored 0.7–1.0; milk and herbs scored near zero. Combining low lift with high humor yields the truly bizarre combos—like condoms and apples—that make the cashier snicker.

The full product-to-brick mapping and humor scores are available as gists. It's a neat example of how a bit of engineering—embeddings, LLM classification, and a custom scoring function—can turn a mundane dataset into something genuinely entertaining.