YouTube Recommendation System

YouTube has tons of videos available on their platform, with thousands of new videos being uploaded every minute. Having this amount of content, it becomes a major task for YouTube to recommend videos to its users which will make sure that a user has great time on the platform. YouTube earns revenue through advertisements, so they want a user to spend maximum amount of time on the platform. So, showing good suggestions according to taste becomes important.

YouTube constantly tries to improve recommendations of a user and his experience. You might have seen something like this while watching videos on YouTube:

Here, I was watching a clip from a test match and right below that YouTube recommended me a video of highlights of The Ashes match considering interest in test cricket and below that they asked how you feel about this suggestion to improve recommendations.

First, let’s understand what are recommendation systems. When you shop online on online stores like Amazon, Flipkart, Myntra, etc. these companies keep track of every click of a user. They track things like which products a user was searching, how much time he spent on a particular product reading details and reviews of product. Suppose, you buy a book of mystery genre, you will see suggestions of books of genres like mystery, suspense, thriller on the site. These suggestions are made possible because of recommendation systems.

Before going in technical part of recommendation system, let’s understand what factors YouTube consider while suggesting videos.

For understanding working of YouTube recommendation system, let’s take sample of few users and content watched by them.

Here, we will take 5 users and refer them as U1,U2,U3,U4,U5.

U1 has watched videos of GTA 5, cricket highlights, hindi songs, funny video. Also, we have similar data of other 4 users. Based on these data, we will categorize those videos. We will name those as Vi (i= 1,2,3,4,5) where V1: Gaming, V2: Cricket, V3: Bollywood, V4: Music, V5: Comedy.

Here we have our data in matrix form:

This matrix tells us which user has watched which types of videos. We can show that in sets as below:

U1: {V1,V2,V4}

U2: {V1,V4,V5}

U3: {V2,V3}

U4: {V2,V3,V4,V5}

U5: {V1,V3,V4,V5}

Obviously, YouTube will recommend videos to an individual according to his taste but also they will try to see similar behavior of users and try to recommend according to that.

We will see similarity of watched videos of two users in our data. We will do this by simply taking intersection of sets of two users. We will get similarities as follows:

So, we get highest similarities for (U2 and U5) and (U4 and U5) which is 3. In this case, even if a video is not according to an individual’s taste but that user follows similar behavior as another user then he/she will get recommendations of those videos which hasn’t been watched by them.

For U4 and U5, both have watched 4 types of videos. U4 has not watched gaming video and U5 has not watched cricket video. Both of them will get those videos in recommendation respectively.

For U2 and U3, similarly is zero. So they might not get recommendation of videos they have not watched but other has watched.

Hope you might have get an idea about how YouTube recommendation system works. Of course, this is just a simple understanding of system. Actual system is much complex.

Thank you…

Related posts:

If you are an e-commerce business or even are related to the e-commerce space, you must have come across the word Headless Commerce. With an API-driven decoupled architecture, there is no doubt that…

I have a couple of friends who have been through a lot, but are now stable with good jobs and spouses. They are at stages of their lives where they have less stress coming in, but they are still…

The pigeonhole principle states that if you have n+1 objects (or pigeons), and n groups (pigeonholes), then one group must have more than one object. This remains true for any m objects > n groups…