It's interesting but they could have given more details into the features, types of models they tried, how successful it is etc etc. right now it just reads as "we used ML to fix a problem we have".
My experience with Netflix's video quality has been pretty positive, so they seem to have been doing something right. I hope this predictive model will only improve things and won't deteriorate quality (consider they tuned for low false negatives quality shouldn't decline).
This is a nice write-up of the business case for data science. It would have been nice to have a follow up with the technical details - but maybe that's deemed proprietary?
My problem with Netflix is that I rarely find streaming films I would be interested in watching. It's so bad that I don't even bother looking there any longer, I'd rather just pay the $2.99 and rent a film I'd like to watch from Play or Amazon.
> My problem with Netflix is that I rarely find streaming films I would be interested in watching
The Netflix TV catalog is very deep. What my wife does is pick a show that has 7+ seasons, and then starts watching it. If she likes it after the first few episodes, then generally it will occupy her for a few months.
Just curious: what are you interested in that Netflix doesn't have? My queue always hovers around 350 unwatched titles, so this always baffles me a little bit.
I discovered this thing the other day, https://www.smartflix.io/ and lo behold, imagine my surprise when nearly everything I searched for... was in fact, streaming in my own country (New Zealand) and nowhere else. That was an eye opener.
With that said, I quite like the Smartflix UI -- it's great for discovering things.
I started enjoying Netflix more when I realized I shouldn't be looking for stuff I knew of from elsewhere, but rather explore titles, rate, and help their algorithm to suggest things to me. Has been working out well so far - I got into an area I didn't think I was interested so far (superhero series).
So much (great) emphasis on viewing experience, so little emphasis on browsing experience. I've been a customer for over a decade and am just perpetually frustrated; e.g. on the Netflix PS3 app titles automatically start playing when you view their info. So if you wanted to watch that title, it saves you the agony of clicking "Play." If you were just browsing, you now have this thing playing while you're trying to read the synopsis, director, actors, etc. Then, it'll show up in Account Activity as "started" and also affects recommendations ("Based on your interest in...").
Have had the same issue on my PS4. It's mildly annoying when searching for something new, and kinda nice when you're getting back into something you were watching yesterday.
I wish it was selective about autoplay, for certain conditions. Chances are, I don't want it to play when I've never watched a second of any episodes in my life, but it's nice that it autoplays when I click it from "Continue Watching" section.
Same thing for the Amazon Fire Stick... the Netflix app is extremely laggy and playback stutters, it feels like even they have very capable engineers, there's a lack of love at the very last mile.
Also, discovery of new films is awful, as they basically want to be helpful and guessing your taste all the way, but it's severely lacking filtering options for power users. Which I don't mind, because it's driving people to our site in droves recently :)
The sheer amount of systems they are trying to cover with a performance and security (DRM) sensitive video player component isn't that easy to support on the other hand.
I don't even think basic filtering is only appealing to power users. Just some way to find films apart from their recommended, too-cute "genres" would be appreciated.
If this kind of control system interests you, check out Chimera [0]. It is a similar approach that the output of the machine learning approach can go to a human, but is interesting because the human can feedback to the machine learner. That is, not only can the human see that the machine learner was wrong and write rules to correct it next time, but the learners can utilize that new training data to better classify in the future as well.
Of course its a tradeoff, because it seems like false negatives (movie had QoE problems but ML system did not flag it) should be avoided at all costs in this case.
15 comments
[ 5.0 ms ] story [ 39.3 ms ] threadNot to say this can't be the exception, just that it does seem a little out of character.
This is a nice write-up of the business case for data science. It would have been nice to have a follow up with the technical details - but maybe that's deemed proprietary?
My problem with Netflix is that I rarely find streaming films I would be interested in watching. It's so bad that I don't even bother looking there any longer, I'd rather just pay the $2.99 and rent a film I'd like to watch from Play or Amazon.
The Netflix TV catalog is very deep. What my wife does is pick a show that has 7+ seasons, and then starts watching it. If she likes it after the first few episodes, then generally it will occupy her for a few months.
With that said, I quite like the Smartflix UI -- it's great for discovering things.
I wish it was selective about autoplay, for certain conditions. Chances are, I don't want it to play when I've never watched a second of any episodes in my life, but it's nice that it autoplays when I click it from "Continue Watching" section.
Also, discovery of new films is awful, as they basically want to be helpful and guessing your taste all the way, but it's severely lacking filtering options for power users. Which I don't mind, because it's driving people to our site in droves recently :)
(shameless self-plug: https://www.justwatch.com )
The sheer amount of systems they are trying to cover with a performance and security (DRM) sensitive video player component isn't that easy to support on the other hand.
Of course its a tradeoff, because it seems like false negatives (movie had QoE problems but ML system did not flag it) should be avoided at all costs in this case.
http://pages.cs.wisc.edu/~anhai/papers/chimera-vldb14.pdf