What's interesting about a by-the-numbers account of Flappy Bird's success is how starkly apparent the weaknesses of the app store model of surfacing and distributing apps seem to be in retrospect. Think about it: the app store organizes apps almost exactly like web directories organized websites in the 90's -- there are human-curated catalogs, and a single store acts as a directory for essentially all known apps. It's like Yahoo! was in 1994.
Considering that Flappy Bird was made popular by power users on other platforms (particularly YouTube), I think that this point is really important: the app store is probably not the best platform to surface new apps.
A good exercise for entrepreneurs is to think of a better model. If anything, the success of Flappy Bird is one hell of an incentive to find out.
(Also: if you have ideas about this, like me, we should have a chat.)
Discovery in general is still a pretty knotty problem.
Even Netflix, which has a relatively small pool of things to sort through, doesn't do a particularly good job of it. It often recommends things I have no interest in whatsoever, and conversely never recommends things I very much want to watch but have to find manually.
It's practical to manually sort as a Netflix customer precisely because their pool is relatively shallow. But I wonder if their weak recommendations are due to that shallow pool, and that in the end there just aren't enough ratings for machine learning algorithms to actually work the way they're supposed to.
There's obviously a lot of money in getting recommendations and discovery to actually work, but I wonder if everyone is on a path that's sort of condemned to failure due to lack of enough data.
Dunno. Thinking out loud really. I know it's something I'm dissatisfied with pretty much across the board.
> Even Netflix ... doesn't do a particularly good job of [discovery].
I see this message repeated by a lot of tech people, but I feel like I get fairly good recommendations. Am I alone?
What is it about the recommendations that seem bad? Is it because you know there is better content there but the metadata has somehow not categorized it properly for cross-referencing, or does it seem more like it's guessing from a shallow pool and trying to bubble up whatever content it can find, even if it's not a great match?
Is there any kind of open source recommendation framework or standard? Couldn't Netflix publish the details of its algorithm, including which metrics it has, to help the community think up more ways to analyze? The value in Netflix right now is in the content, not in the recommendation engine.
Looks like the competition Netflix used to host was ended in 2009... would be great if they started it again: http://www.netflixprize.com
I had a response written to this, but realized I haven't thought it through enough to respond meaningfully.
I know I'm dissatisfied. I know it doesn't show me movies I very much want to see, and that it clutters my recommendations with a lot of things I'll never watch.
My feeling is that it doesn't capture why I like the movies I like.
The categories they've been refining over the years are likely an attempt to address that, but they seem to be missing the important information still.
Hmm.
PS- re the contest, there was a problem with the anonymized data being de-anonymized by researchers, so they couldn't make it available any more.
I suspect it's that last one. I mostly use netflix in streaming mode. The pool of available content netflix is allowed to stream at any given time is pathetic - most of the best matches they could make are films they don't have so they have to settle for suggesting something half-assed that they do have.
As a result, I find myself gradually tending more and more to use various other streaming services first, and only settle for Netflix if I can't find it somewhere else first. (I should just cancel Netflix entirely but haven't gotten around to it yet.)
Their suggestions seemed better when they were a mail-only service and hence not subject to the same limited pool. (A separate factor: when I initially got Netflix there was a HUGE list of movies I wanted to see and hadn't yet; over time that pool also has substantially diminished. So there might be fewer good matches to find even if Netflix had unlimited scope to provide the good matches.)
I briefly used Netflix several years ago, and was astonished at how good the recommendations were. However, I aggressively rated many movies in their catalogue to get better recommendations, movies I'd seen before not necessarily just movies I'd watched on Netflix. When I talked to others who had it, and didn't like their recommendations, it seemed to be because they weren't rating the movies at all.
Maybe they couldn't be bothered, maybe they assumed it would know based on what they watched, or maybe because these were couples I often talked to it was awkward to bring up rating the movie when there were potentially two different opinions on it, and one or both parties unwilling to debate the issue...
Given my anecdotal experiences of this problem I would guess that Netflix's problem isn't a matter of figuring out what I would like, it's a matter of figuring out what I may not have seen yet. The problem is that the things that I am most likely to like based on my viewings are also the things that I am most likely to have already seen somewhere other than Netflix.
I'd argue that Netflix is high-precision, low-recall: the quality of recommendations are usually good, but the recommendations themselves usually don't reflect the entire library.
Great point. I've been considering working on a tool that tried to solve this problem for health apps. It would require the user to fill out a form that teases out information about their motivations, health goals, lifestyle, etc.(the more data the better the recommendation). They'd then receive an auto generated menu of apps that'd work specifically well for them.
I've spent a lot of time talking to my Dad about healthcare (he's a doctor) and he never recommends mobile apps to his patients for weight loss or anything along those lines. I can't really blame him... there are tens of thousands of health apps. That might change if he had an online tool to send his patients to that did the heavy lifting of making an intelligent recommendation. Here's an interesting study that shows how doctors can be a powerful catalyst in this regard: http://www.thelancet.com/journals/lancet/article/PIIS0140-67...
In the end, it'd just be an app recommendation engine for health. That aside, I 100% agree with what you're saying.
Google play annotates and highlights apps that you or your friends have downloaded or rated. 'similar apps' is also a more modern way of surfacing content.
Sure, but how do you deal with the long tail? Flappy Bird is not an app that would have surfaced until it was well on its way to becoming what it is today (or was yesterday, I guess).
Considering that Flappy Bird was made popular by power users on other platforms (particularly YouTube), I think that this point is really important: the app store is probably not the best platform to surface new apps.
A good exercise for entrepreneurs is to think of a better model. If anything, the success of Flappy Bird is one hell of an incentive to find out.
(Also: if you have ideas about this, like me, we should have a chat.)