Spotify rediscovers what others found a decade ago, social recommendations don't work, that "no matter who you are, someone you don't know has found the coolest stuff." (for more on that, see http://longtail.typepad.com/the_long_tail/2005/02/why_social_netw.html)
6 plus ones
Shared publicly•View activity
- HT,Dec 7, 2012
- That's a bit harsh. You are one of my news source. I absolutely never rely on automated or collaborative news recommendations... I very carefully select a few people that I follow. (For example, you might have noticed that I react to a lot of what you write... I can afford to since I don't follow lots of people.) I do the same thing in science with research... I tend to follow a few key researchers...
In theory, I like Amazon's recommender system, but in practice it is thoroughly useless now for me.
Disclaimer: I am someone who worked on recommender systems in the past.
I have not yet formalized the reason behind this strategy I follow... but it goes like this... an automated or collaborative system provides recommendations that have no "meaning". They end up stressing me out (there is so much out there). By carefully "following people", I end up with a better experience... because, I suppose, that I feel a connection to you... so the news article are not just random stuff that I might be interested in... they have meaning because they come from you.
I guess it is related to the following concept. A given drawing has probably no value to you... but what if your kid drew it? Then it might become the most precious thing in your office!Dec 7, 2012
- A carefully curated list of news sources can be very useful, agree, but it takes a lot of work to maintain. The advantages of implicit recommendations are that it requires little effort and is anonymous.
But I think you're right that I may be too harsh here. And, in particular, I think the way I stated it implies a wide gulf between recommendations and social, but I really see them as flavors of the same thing. Reading something because an expert recommended it, reading something because it's popular, reading something because a friend recommended it, reading something because someone you don't know with similar tastes liked it, these are all flavors of discovery, all sources of data for discovery, all ways of finding things you hadn't found yet.
Recommendations work by finding experts and people with similar tastes you don't know about and telling you what they found, so it's very similar to a curated list, just without the work (but also without the control) of curating the list yourself. Really is a flavor of the same thing, with some benefits (less work and anonymous), but some costs (less control, possibly lower quality depending on the effort you're willing to put in to curating your list carefully and filtering stuff yourself).
By the way, I agree the Amazon recommender system has gotten less useful. I had a post about that a while back, that Amazon seems to have squandered its early lead in personalization, here it is: https://plus.google.com/u/0/102076128417589427747/posts/Ed8PBUbFfzSDec 7, 2012
- Yes. But I am not claiming that Amazon's recommender system is bad from a Computer Science point of view. I'm sure it is quite sophisticated. It is just not how I find stuff that I like. I wonder whether I'm an exception... but I just view their recommendations as noise... I'd rather not have it on the page. BTW, is it just me or is Amazon's main page getting more crammed with junk every year? It was always a bit bad... but right now it is just plain ugly.Dec 7, 2012
- Completely agree on Amazon getting more crammed with junk. They've been doing a lot of hiring, including a lot of PMs out of Microsoft, and, like Microsoft, I suspect a lot of those PMs are rewarded for delivering features (launching their feature, regardless of how useful it is or how well it performs). Lots of feature creep is the result.
For a while now, I've been thinking Amazon needs a group of fearless people who just runs A/B tests of improving the stuff from other groups (which, in many cases, might mean removing features entirely). I bet some simpler detail and home page designs, for example, that remove a bunch of mostly distracting features could win A/B tests by a wide margin. Wouldn't have to roll out removing the feature, necessarily, just knowing how much keeping the feature up as is costs the company would be valuable. Not sure people in that group would be popular, but it sure would be fun.Dec 7, 2012
- Wow, just incredible how all these issues just go in loop.
Regarding the following - friends indeed have bad taste in music, news, gadgets, books etc. It is similar to the 80-20 principle - 80% of the users use 20% of the features, but each other - a separate 20% set.
But, I still believe that usage-based recommendation system - this one may be automated. If me and someone else "like" the same items, then we can be expected to "like" the same items in the future and then the "items" I like can be proposed to the people who liked similar to me in the past.
The issue here is just how to cut the "like" into subsets, both thematically and in time.
As for me, I rely on maintained news source, but there is no normal app to do it in a centralized way, and automate "promote" and "relegate" for my sourcesDec 7, 2012
- Doesn't Amazon require lots of A/B testing before launching all of those features in the first place?Dec 8, 2012
- PMs who are rewarded on launching features (whether or not they work) have a strong incentive to skip or game the A/B tests, which is what happens.Dec 8, 2012
Add a comment...