9 ms·

I face this problem with Netflix, Spotify and Youtube too. These algorithmic recommendations just need one improper dataset to throw everything out of the windo

by inapis 7y ago

I face this problem with Netflix, Spotify and Youtube too. These algorithmic recommendations just need one improper dataset to throw everything out of the window. For weeks, my spotify is overloaded with instrumental songs. Youtube keeps on repeating the same stuff. Netflix believes that the only thing I watch is science fiction. I'm dying for human curation.

rhizome 7y ago
You also have to account for payola.
- xbkingx 7y ago
  Which is why the author of the article didn't find the upcoming release for the book she searched. Searching for a book we can't make money on? Pffft GTFO.
icebraining 7y ago
Criticker is automated, and I find its recommendations amazing. (Not affiliated, just a happy user)
- nefitty 7y ago
  Criticker has never failed me. I've been using it for years. I did hop over for a few years to a site called Jinni with really high quality UI. They closed their consumer facing services and provide their service to content providers. Criticker is still there crushing the game.
chewxy 7y ago
That's because the recommendation systems are optimizing for the wrong things: typically dwell time. You are highly likely to dwell longer on scifi, so heres more scifi for you! Older recommendation systems use background and foreground timeframes to build frames of references, hence repeats. The question to ask then is: what is the correct thing to optimize for? Dwell time is usually chosen for advertising and stickiness (you are more likely to stick with a service that plays what you like). Optimizing for novelty is difficult because the set of unknown things is much larger than the set of known things. Plus, it is risky from a business point of view.
- kristopolous 7y ago
  It's an interface problem. Users aren't able to express the fidelity needed through the current action set
  - mooreds 7y ago
    Right. You only have four actions on Spotify as I recall (thumbs up, thumbs down, pause and fast forward). I guess some of the other apps have the larger bucket of 'who is watching/acting' based on account. I am not sure how many actions are needed, but I can think of a few more: * Not now, but later is fine * Maybe next month * Play the opposite of this * Something similar but new to me
    - ancarda 7y ago
      It'd be nice to have an "I've already seen this" button that will mark an episode/season/show as watched so it doesn't keep coming back up. Netflix really wants me to watch Breaking Bad, but I've already seen it. I liked it; give me more like that, but not exactly that
      - frosted-flakes 7y ago
        You can do this with YouTube. Just click "Not interested" in the three-dots menu, and click the "Tell us why" link that appears.
        wwwhizz 7y ago
        After which it will still recommend it to you.
      - DJHenk 7y ago
        > It'd be nice to have an "I've already seen this" button that will mark an episode/season/show as watched so it doesn't keep coming back up. That's not just nice, that is the most obvious "feature" a normal person would design. I've stopped watching Netflix because of nonsense like this. They know I've seen it before, because I watched in their player. Also, even the tiniest, most modest application of analytics would discover that I always watch one series at the time, from start till end, or whenever I'm bored with it. In the past years, I have never ever watched a thing twice. Still, it was constantly showing me things I've already seen. The only conclusion I can draw from this is that Netflix does not have have my best interest in mind when designing their algorithms and that they don't respect their customers. So I canceled.
ztarven 7y ago
I absolutely agree, algorithms are terrible for recommendations. On the media database[0] my wife and I are building we're experimenting with community upvotes for suggestions. Our userbase is small so the results are inconclusive right now but hopefully time will tell if this method works. [0] https://rate.house https://rate.house
- inimino 7y ago
  There's no reason algorithms have to be terrible, after all, most of what we do is expressible algorithmically when we give recommendations (similar interests, right level for the reader, not too advanced or too introductory, something they can use or that leads to other insights, etc.) It's just that the algorithms are generally so bad because they are optimizing for the wrong things and not any of these things that actually make for good recommendations.
- bscphil 7y ago
  The only good one I've ever seen was in the music business: Rdio, specifically. (Now sadly dead.) I don't know what they were doing, but their recommendations were consistently excellent at pushing me towards new artists and even new genres that I didn't expect but would up enjoying. Back when I used to use it (about a decade ago), Pandora was good at shuffling, but their library was so small I would get the same songs over and over.
WalterBright 7y ago
It'd be nice to simply tag items with keywords. Then you can search for matches with various boolean combinations of keywords.
xbkingx 7y ago
I call it "collapsing into the mean" - all recommendation engines (as the currently exist) will eventually corral you into the most vanilla, mass marketed set of recommendations and then fail miserably when any conflicting data is presented. We've effectively turned the web into cable TV circa 1990 - a finite set of junk food level entertainment sources that we voted for because they were "eh, good enough" and easy to find. With as many YouTube videos that I've watched since 2005, you'd think they might recommend a video with less than 1000 views once in a while. I've found 1 new channel in the last 6 months and used the "Not Interested" option more times than I can count. And the rules are unclear. I don't want tech reviews from 5 years ago, but if I hit Not Interested, does that influence the channel, topic, age, keyword, etc recommendation frequency? I don't mind old DIY or woodworking videos, I'm subscribed to the channel, and watched the recommended video 5 years ago. (Of the current 12 YT Recommended videos, 2 are labelled Watched and only 2 are less than a year old.) I think part of it stems from the lack of user organization features. Offer too many and you scare away users, while the people with vested interests dump money and time in to wash out any negative opinion (see Amazon reviews). Offer too few and you get poor recommendations during on-boarding/startup.
- maest 7y ago
  > And the rules are unclear. There are no rules. They just record your preferences and might retrain their black box algo in the future. The black box decides those rules
technofiend 7y ago
They often don't expose knobs I'd really like to tweak, either. For instance if I grew up listening to original album version of something I don't want to hear the live one. But there's only an upvote/downvote choice on Google music so what exactly am I giving the thumbs down? The artist? The song? This particular version of the song? No idea.
hinkley 7y ago
Why for the love of god and your own mothers would they suggest movies to me that are similar to a movie I watched 5 minutes of and never finished? If I can't make it through 5 minutes of a movie I sure as fuck don't want to watch another movie that's just like it. Read the room already.
- perl4ever 7y ago
  I've found it odd that Spotify keeps playing me the same tracks even if I forward past them every single time they start because I can't stand it. I don't think that can be stupidity, it must be malice...
jillesvangurp 7y ago
Spotify has one genious feature that I use a lot: similar artists. The problem with machine learning and recommendations is that it has a few pitfalls that lead to poor results. This is on full display on Amazon which manages to recommend stuff I've already bought from them while not being able to tell the difference between hard science fiction and fantasy. The resulting recommendation bubble seems impossible to escape. With Spotify, they do have a few features that work for me. I've discovered new artists by exploring their "Fans Also Like" feature. The nice thing about this is that they don't try to be too smart there. Their normal recommendations suffer from the same issues that other sites have and are thrown off by the fact that my tastes are all over the place. I happily listen to sixties psychedelic rock, jazz, metal and some techno or some punk and tend to go from one to the other. Yet I'm very picky about what I listen to. Somewhere along the lines it seems to have decided I'm a middle aged guy (correct) and it consistently does not recommend me any music made this century; which is kind of frustrating if you are trying to find something new to listen to. Recommendation bubbles are a thing and escaping from them is hard. I work around it by using the fans also like feature and using it's suggested additions to playlists. This works surprisingly well. Example based similarity search is a much simpler problem then recommendations. And it's IMHO a much more interesting feature to explore content with.
- romack77 7y ago
  The desire for these recommendation services not to be "too smart" really resonates with me. If we can't yet give a great set of zero-effort recommendations, why don't we pull back a step and give the user a few power tools to find their own? Maybe I'm out of step with mainstream users, but I would see that as a huge improvement.
rwmurrayVT 7y ago
You watch one little anime with someone for high school nostalgia and suddenly Netflix thinks I'm only interested in k-pop docs, anime, and unusual shows featuring all-Asian casts... ???

by inapis 7y ago

rhizome 7y ago

You also have to account for payola.

xbkingx 7y ago
Which is why the author of the article didn't find the upcoming release for the book she searched. Searching for a book we can't make money on? Pffft GTFO.

icebraining 7y ago

Criticker is automated, and I find its recommendations amazing. (Not affiliated, just a happy user)

nefitty 7y ago
Criticker has never failed me. I've been using it for years. I did hop over for a few years to a site called Jinni with really high quality UI. They closed their consumer facing services and provide their service to content providers. Criticker is still there crushing the game.

chewxy 7y ago

That's because the recommendation systems are optimizing for the wrong things: typically dwell time. You are highly likely to dwell longer on scifi, so heres more scifi for you! Older recommendation systems use background and foreground timeframes to build frames of references, hence repeats. The question to ask then is: what is the correct thing to optimize for? Dwell time is usually chosen for advertising and stickiness (you are more likely to stick with a service that plays what you like). Optimizing for novelty is difficult because the set of unknown things is much larger than the set of known things. Plus, it is risky from a business point of view.

kristopolous 7y ago
It's an interface problem. Users aren't able to express the fidelity needed through the current action set
- mooreds 7y ago
  Right. You only have four actions on Spotify as I recall (thumbs up, thumbs down, pause and fast forward). I guess some of the other apps have the larger bucket of 'who is watching/acting' based on account. I am not sure how many actions are needed, but I can think of a few more: * Not now, but later is fine * Maybe next month * Play the opposite of this * Something similar but new to me
  - ancarda 7y ago
    It'd be nice to have an "I've already seen this" button that will mark an episode/season/show as watched so it doesn't keep coming back up. Netflix really wants me to watch Breaking Bad, but I've already seen it. I liked it; give me more like that, but not exactly that
    - frosted-flakes 7y ago
      You can do this with YouTube. Just click "Not interested" in the three-dots menu, and click the "Tell us why" link that appears.
      - wwwhizz 7y ago
        After which it will still recommend it to you.
    - DJHenk 7y ago
      > It'd be nice to have an "I've already seen this" button that will mark an episode/season/show as watched so it doesn't keep coming back up. That's not just nice, that is the most obvious "feature" a normal person would design. I've stopped watching Netflix because of nonsense like this. They know I've seen it before, because I watched in their player. Also, even the tiniest, most modest application of analytics would discover that I always watch one series at the time, from start till end, or whenever I'm bored with it. In the past years, I have never ever watched a thing twice. Still, it was constantly showing me things I've already seen. The only conclusion I can draw from this is that Netflix does not have have my best interest in mind when designing their algorithms and that they don't respect their customers. So I canceled.

ztarven 7y ago

I absolutely agree, algorithms are terrible for recommendations. On the media database[0] my wife and I are building we're experimenting with community upvotes for suggestions. Our userbase is small so the results are inconclusive right now but hopefully time will tell if this method works. [0] https://rate.house https://rate.house

inimino 7y ago
There's no reason algorithms have to be terrible, after all, most of what we do is expressible algorithmically when we give recommendations (similar interests, right level for the reader, not too advanced or too introductory, something they can use or that leads to other insights, etc.) It's just that the algorithms are generally so bad because they are optimizing for the wrong things and not any of these things that actually make for good recommendations.
bscphil 7y ago
The only good one I've ever seen was in the music business: Rdio, specifically. (Now sadly dead.) I don't know what they were doing, but their recommendations were consistently excellent at pushing me towards new artists and even new genres that I didn't expect but would up enjoying. Back when I used to use it (about a decade ago), Pandora was good at shuffling, but their library was so small I would get the same songs over and over.

WalterBright 7y ago

It'd be nice to simply tag items with keywords. Then you can search for matches with various boolean combinations of keywords.

xbkingx 7y ago

I call it "collapsing into the mean" - all recommendation engines (as the currently exist) will eventually corral you into the most vanilla, mass marketed set of recommendations and then fail miserably when any conflicting data is presented. We've effectively turned the web into cable TV circa 1990 - a finite set of junk food level entertainment sources that we voted for because they were "eh, good enough" and easy to find. With as many YouTube videos that I've watched since 2005, you'd think they might recommend a video with less than 1000 views once in a while. I've found 1 new channel in the last 6 months and used the "Not Interested" option more times than I can count. And the rules are unclear. I don't want tech reviews from 5 years ago, but if I hit Not Interested, does that influence the channel, topic, age, keyword, etc recommendation frequency? I don't mind old DIY or woodworking videos, I'm subscribed to the channel, and watched the recommended video 5 years ago. (Of the current 12 YT Recommended videos, 2 are labelled Watched and only 2 are less than a year old.) I think part of it stems from the lack of user organization features. Offer too many and you scare away users, while the people with vested interests dump money and time in to wash out any negative opinion (see Amazon reviews). Offer too few and you get poor recommendations during on-boarding/startup.

maest 7y ago
> And the rules are unclear. There are no rules. They just record your preferences and might retrain their black box algo in the future. The black box decides those rules

technofiend 7y ago

They often don't expose knobs I'd really like to tweak, either. For instance if I grew up listening to original album version of something I don't want to hear the live one. But there's only an upvote/downvote choice on Google music so what exactly am I giving the thumbs down? The artist? The song? This particular version of the song? No idea.

hinkley 7y ago

Why for the love of god and your own mothers would they suggest movies to me that are similar to a movie I watched 5 minutes of and never finished? If I can't make it through 5 minutes of a movie I sure as fuck don't want to watch another movie that's just like it. Read the room already.

perl4ever 7y ago
I've found it odd that Spotify keeps playing me the same tracks even if I forward past them every single time they start because I can't stand it. I don't think that can be stupidity, it must be malice...

jillesvangurp 7y ago

Spotify has one genious feature that I use a lot: similar artists. The problem with machine learning and recommendations is that it has a few pitfalls that lead to poor results. This is on full display on Amazon which manages to recommend stuff I've already bought from them while not being able to tell the difference between hard science fiction and fantasy. The resulting recommendation bubble seems impossible to escape. With Spotify, they do have a few features that work for me. I've discovered new artists by exploring their "Fans Also Like" feature. The nice thing about this is that they don't try to be too smart there. Their normal recommendations suffer from the same issues that other sites have and are thrown off by the fact that my tastes are all over the place. I happily listen to sixties psychedelic rock, jazz, metal and some techno or some punk and tend to go from one to the other. Yet I'm very picky about what I listen to. Somewhere along the lines it seems to have decided I'm a middle aged guy (correct) and it consistently does not recommend me any music made this century; which is kind of frustrating if you are trying to find something new to listen to. Recommendation bubbles are a thing and escaping from them is hard. I work around it by using the fans also like feature and using it's suggested additions to playlists. This works surprisingly well. Example based similarity search is a much simpler problem then recommendations. And it's IMHO a much more interesting feature to explore content with.

romack77 7y ago
The desire for these recommendation services not to be "too smart" really resonates with me. If we can't yet give a great set of zero-effort recommendations, why don't we pull back a step and give the user a few power tools to find their own? Maybe I'm out of step with mainstream users, but I would see that as a huge improvement.

rwmurrayVT 7y ago

You watch one little anime with someone for high school nostalgia and suddenly Netflix thinks I'm only interested in k-pop docs, anime, and unusual shows featuring all-Asian casts... ???