Hacker News Reader

Comment by pcbro141 | Hacker News Reader

Why doesn't Amazon have amazing or even just 'good' product recommendations if they have all these ML experts?

I think it’s just a much harder problem than Netflix or Youtube recommendations - Amazon has at least 500 million products, and people aren’t buying individual products at the same scale as they watch videos. Has anyone with >1000 products really nailed product recommendations without large precision losses?

That's so true. And it's not just the relevance part of it, we have to even incorporate few operational cues as well while recommending products. Not for Amazon but for a different marketplace, I manage the recommendation engine as a product and we have incorporated product serviceability to customer's address as one of the parameters. Then the usual e-comm stuff kicks in like, out of stock issues, price elasticity during sale periods, new product/category launches etc.

Obviously, I do not want to discount other non-commerce recommendation engines like Spotfiy's or Netflix. They are great in their own ways, but e-comm is quite a different ball game. The cross sell model is really hard to crack if you have 1000s of competing sub categories and millions of products in them to fill in your top 10-20 slots.

Maybe what you consider "good" recommendations don't make as much money. They extensively A/B test everything and what is in prod is what works.

Recommendation systems are based on aggregated preferences. The farther you are from an aggregate/group or the rarer your group is, the worse your recommendations. It's an inherent limitation. It's in essence trying to guess at missing data by comparing your past purchase patterns to known data. There's also a bootstrapping problem -- if you haven't bought enough stuff to establish a pattern, the algorithm can't really figure out what you'd want next.

If you order a Rode microphone and a Sony RX cam, it might figure out that you're a Youtuber and might recommend a Zhiyun gimbal. The cluster of data points point to that.

But if you have order pattern where the data is sparse, it's not likely to come up with a good recommendation.

Apparently they weren't required to actually fix the problem, just play with ML a bit and push it to production.

Also, ML isn't magic. It's almost as much work as not using ML, if you want actually good results.

Reply on news.ycombinator.com