4 ms·
Tuning relevance is extremely difficult, and it’s not just an information retrieval problem - at Amazon’s size (and even smaller companies) a lot of what you se
by lukevp 6y ago
Tuning relevance is extremely difficult, and it’s not just an information retrieval problem - at Amazon’s size (and even smaller companies) a lot of what you see for “relevance” is hand-tuned by humans, through machine learning on longer-tail queries, and some combination of both in the middle (with sponsored or high margin or velocity items weighting more heavily). If you search for something like “men’s shoes” there’s 0% chance you’re getting a search result, it’s going to be a curated result page. Lots of people work on this hand tuning, but they’re not engineers, they’re product and marketing people.
If you search for “men’s shoes with green laces” you’ll see the real challenge, with a top 10 search result that’s not even shoes (it’s a pack of laces).
- saalweachter 6y agoSome of my favorite test queries for shopping sites (from back when I used to work on a comparison shopping engine: [apples to apples] -> frequently returns random "Apple" products, because when you dedupe, stem and remove stopwords from that query, you're left with "apple", and Apple products are big ticket items. [shirt dress] -> frequently returns dress shirts, because many sites essentially ignore word order or the notion of compound words. [fish swim vest] -> frequently returns fishing vests, partially because of overstemming (many POS stemmers will stem "fishing" -> "fish", but in a product search setting, it's usually the wrong thing to do).
- H8crilA 6y agoHere's one that's fun if your objects of interest include cities: https://en.m.wikipedia.org/wiki/Best,_Netherlands https://en.m.wikipedia.org/wiki/Best,_Netherlands What's the [best city in europe]? Obviously "Best, Netherlands".