Okay, that makes sense to me. They go from 0.02% likelihood of wanting to buy a fridge to, let's say, 2% likelihood. A marvelous leap!
But even though they are now much more likely to buy another fridge than the average person ... I would think that they would be more likely still to buy other types of products than to buy another refrigerator.
Even if their likelihood of needing to buy more consumable products (e.g. deodorant) is only 5%, that's still double the likelihood of their likelihood to buy a fridge again.
So I would think that there would be some bias toward more frequently purchased items, even if your likelihood of buying a less frequently purchased item does indeed increase after purchase.
I'm sure the math checks out somehow, and that the creators of the recommendations algorithm wouldn't be pursuing strategies that don't work, but I don't think Patrick's explanation fully captures it.
My only explanations are that that kind of thing happens when they can't figure out anything good to recommend.
You seem to be taking this an an axiom and running with it. My impression too, is that the recommendations are pretty shoddy.
- yes, people who just bought a lot of batteries are actually more likely to buy more batteries soon. same for people who just bought a fridge.
- these recommendations may cause some sort of "burn out", where people stop looking, responding, or decide to unsubscribe. But this wasn't much worse than other recommendations. including the awesome ones.
- the loss from burnout was greatly outpaced by the wins in actually targeting the right users
Yes, these are bottom-of-the-barrel recommendations. But they work. And people don't care that much even if they disagree. Maths said so.
I think this is a similar story to the complains on non-literal searches in search engines. People think they want one thing, but they want something else.
Perhaps they are not.
Perhaps the best they can do, is to recommend to all of us the same/similiar product again because thats the only thing they got working: Catching those x percent of people who send the tv back and buy the second recommended one.
Patrick explained it why it might be like so: https://mobile.twitter.com/patio11/status/982208307057246209 (thread: https://threadreaderapp.com/thread/982208307057246209.html )
> I don't remember ever looking at their recommendations and finding them useful.
Amazon does various forms of recommendations (no way limited to just recent purchases) based on
1. Purchases
2. Searches
3. Browsing history (via trackers, ads, affiliate links)
4. Activities (on their digital devices, tagging things on their websites in wishlist, adding preferences in your account etc)
5. Global and Local trends (people in your area/country buy..., people with similar buying pattern buy..., It's Eid and you're Muslim, so you buy...)
...And probably many other signals I might be missing.
They must have got pretty good people working with this and enough data to get their models right.
Also, Amazon makes decisions on new retail businesses to start depending on those signals. https://techcrunch.com/2016/11/03/amazons-private-label-bran...
That, and considering the fact that Amazon has an unshakeable culture of making data-driven decisions for everything it does...
I guess what I'm trying to say is, Amazon and its subsidiaries may not be as dumb at AI/ML as it might seem to be by gauging against one datapoint.
Also see: https://www.cnbc.com/2014/04/09/big-data-knows-youre-pregnan...