• 3 min read
3.2M Instacart Orders Reveal the Weirdest Grocery Combos
Using Instacart’s 3.2 million-order dataset, Roger Dickey ranked the rarest and funniest grocery item combinations from condoms to enemas.
Image: Hacker News
Roger Dickey took a dataset built for recommendation algorithms and used it for the opposite purpose: finding the strangest item combinations in 3,214,874 Instacart orders. The data, originally released by Instacart for a machine learning competition and still available on Kaggle, covers 49,688 unique products, 134 aisles, and roughly 10 products per order on average.
A simple search for the least common product pairs, triples, and quads did not produce especially funny results. Dickey found that the combinatorics swamp the signal: there are about 1.2 billion possible product pairs, roughly 97% never appear at all, and about 22 million occur exactly once. Ranking by raw rarity was essentially useless.
Instead, he switched to lift — the ratio between how often a combination appears and how often it would be expected to appear. That surfaced patterns like Banana and Bag of Organic Bananas, where near-duplicate products make co-occurrence unusually rare. The problem, Dickey argues, is that Instacart’s product catalog is too granular for humor.
Classifying 50,000 products into GPC bricks
To clean that up, Dickey mapped products into GS1 Global Product Classification (GPC) categories. Rather than work with nearly 50,000 products or the too-broad 134 aisles, he grouped them into about 1,000 grocery-relevant “bricks” from a possible 1,697 grocery bricks across 9 relevant segments.
Recommended reading
Google’s Gemini now meters usage by compute, not prompts
He used a two-stage pipeline:
- Generate embeddings for every product and brick with qwen3-embedding:8b
- Shortlist the 10 nearest bricks by cosine similarity
- Use gpt-4.1-mini to pick the best brick for each product
According to Dickey, the classification run cost about $5 and finished in a few minutes with 60 parallel requests.
That produced cleaner low-lift combinations such as Wine – Still with Grains/Cereal – Not Ready to Eat (Frozen), and Beer with Baby Leaves. Still rare, but not necessarily funny.
A “humor index” for grocery shopping
So Dickey added another layer: a 0–1 humor score for roughly 1,000 grocery bricks, generated by Claude and weighted toward taboo or comedy-adjacent products. Everyday categories like Milk and Herbs scored near 0.0–0.1, while Condoms, Intimate Lubricants, Contraception, and Enemas/Douches landed in the 0.7–1.0 range.
Combining rarity with that humor score produced much better results. Among the standout rare pairs were:
- Garlic + Diarrhoea Remedies
- Kale + Enemas/Douches
- Flat Parsley + Condoms
- Baby Food + Adult Diapers
Triples included Cheese, Almond Milk, and Intimate Lubricants, while one quad paired Cheese, Milk, Apples, and Condoms.
Dickey also narrowed the search to very small carts, where the whole order is only 2 or 3 items. That turned up combinations like Vitamin D Milk and Ultra Thin Condoms, Oreo Chocolate Sandwich Cookies and Personal Lubricant, plus three-item carts such as String Cheese, Black Cherry Yogurt, and Personal Lubricant.
The result is part data experiment, part comedy engine — and a reminder that the same retail data used to optimize recommendations can be repurposed to surface the baskets no recommendation system would ever want to predict.
Computing Editor
Tomas lives in the terminal. He covers chips, laptops, and operating systems with a focus on performance and efficiency. He reads kernel changelogs the way other people read fiction, and he's always on the hunt for the perfect mechanical keyboard switch. If it processes data, Tomas has an opinion on it.
via Hacker News
