Various desktop/ mobile applications (say Instagram, Netflix, Youtube, etc) give recommended content to a user.
How are they able to do that? I wanted to learn what’s the technical aspect that goes behind it.
In: 4
you remember those cookies you always have to accept when browsing for stuff, well those are kind of like trackers that check what you browsed for. then companies such as the ones you listed buy that info from whatever site you browsed in the first place to then show you ads for whatever you browsed for.
Math/algorithms
They look at what content you watch/subscribe/like and then they see what other content do people watch that also watched those contents, and see the engagement level (did they watch all the way, did they comment, etc.). As well as looking at other things.
Algorithms.
Basically a set of instructions for a computer to follow and learn from.
In the case of recommendations on Netflix, the Netflix servers are processing all the watching data that people are doing, finding similarities in the data, and using the similarities to make predictions for other people.
Example: 5000 people watch Show A. Out of the same 5000 people, 4000 also watch Show B.
Algorithm decides that these shows must have a common appeal to have such a big viewer overlap.
So if new person watches Show A, they will subsequently be recommended Show B by Netflix, with the algorithm logic being ‘if 4 out of 5 people watching Show A liked Show B, this person will probably like it too’
Every time you make the choice to watch something on Netflix, your decisions are logged as part of your user data, and the algorithms can make more accurate predictions (in theory) with more information.
It’s the same principle with everything that recommends you stuff based on your consumption. Open a new Amazon account and buy ten Stephen King novels and you will be recommended more Stephen King novels… And also things that other people who have bought a lot of Stephen King novels have bought.
Cookies is one part. Sneaking/watching your pc is another. They use all this data and funnel it into an algorithm called “Recommender systems” (or recommendation systems).
There are three general approaches to it:
1. Collaborative filtering
2. Content-based filtering
3. Hybrid
For t#1, the general idea is that people who liked the same things in the past, would also like the same thing in the future. Those “people who liked XYZ also liked ABC” types of recommendations are usually collaborative filtering. For example, yt may realize that people who liked content creator XYZ in the past, tend to also like channel ABC. So if you watched XYZ, yt may suggest to you ABC.
For #2, the focus is on key features of the item you’re recommending. For songs this is genre, artist, bpm, intensity, volume (and like 15 other audio metrics), etc. And then the program (Spotify) builds a user profile. Based on this profile, the algorithm learns you like songs within the rock genre, at a specific bpm, a specific intensity, volume, etc. And then it will suggest similar songs/artists within the range of that.
#3 is simply a hybrid approach of #1 and #2.
It’s important to note that these algorithms are also learning. The more you interact with them, the more fine-tuned your personalized recommendations would be (to a degree). So if you were recommended band ABC after listening to band XYZ, and you didn’t like ABC (either by ignoring it completely, stopping after a few seconds, or choosing the “don’t recommend me” option if available), the algorithm learns and tries to suggest to you band DEF, which is still somewhat similar to XYZ but with the added info that you don’t like ABC. Etc etc.
But as to what technology yt/fb/IG actually uses, often that is a bit proprietary, and often extremely complex. The approaches above are just very basic and general approaches to recommender systems.
in principal it is really quite simple. When someone clicks on a video, watch what video they go to next, and which video they came from. keep a count of this for everyone
when someone else clicks on the same first video, if they haven’t seen the video most people go to/came from, recommend that to them.
In practice, its quite complicated since you want to filter out noise from the system, and not everyone is the same. So you try to use an “algorithm” to build up a profile of each individual, and what they like/dont like, and what people like them like/dont like. Then, recommend them things people like them have liked (feel free to pick whatever metric you want for what “liked” means, just know that whatever you pick, that content will do better on your platform) (algorithm too, and when in doubt throw marketing keywords at it like “big data”, “machine learning” or “AI Driven”).
From a technical “how do you get the data”, you are logged into these sites, they have your entire watch history and how far you got in each one (so you can continue watching it) all the data is there for the taking. a lot of people here talk about the tracking cookies, and they do uses those for people without an account, but once you are logged in, the tracking cookie is basically useless (unless you go off site).
Latest Answers