How Do Algorithms Work On Social Media Platforms?

Recommendation systems on social media are today’s most mysterious, but also the most rewarding secret. Those who begin to understand the algorithms, capture trends and a large audience. Each social network has its own recommendation systems and its own nuances, and the information on the Internet is very different, to the point that even the developers themselves do not fully understand how their algorithms work.

One of the most striking examples of recent years is TikTok. Thanks to its accurate recommendation algorithms, the short video service has managed not only to overtake all competitors, but also to force them to copy the entire system. It can be said that TikTok algorithms have turned the content of social networks upside down.

Technical types of algorithms

Recommendation systems can vary greatly from each other and use different data. Before we analyze the recommendation systems of individual social networks, let us consider the technical types of algorithms used to create them.

Collaborative filtering

The collaboration system is based on the fact that if users had similar interests in the past, their interests will be similar in the future.

User-based

The scheme is simple: two users A and B have similar preferences in music and performers. If user A liked a song that user B has not yet heard, then it is likely that the listener B will also like it.

Such a principle is based on statistics about user preferences.

Item-based

A similar principle from the Collaborative filtering. In this case, the principle is based not on user preferences, but on the similarity of the object itself.

For example, users usually listen to songs A and B. If a person starts to like song A, they are offered to listen to song B.

Content-based

The essence of this principle is that for each user a separate “virtual portfolio” is created, which takes into account the characteristics of the elements (style, year, author, etc.).

Such a “portfolio” is created based on the user’s preferences, or is asked directly.

For example, a person listens to punk rock and metal by certain artists, which means that they will be recommended similar styles and authors.

Knowledge-based

The system cannot rely on the history of purchases, so it collects data at the very beginning.

For example, a person wants to buy a TV. Usually, we buy a TV every few years. If the system recommends TVs based on the preferences of other users, there are two risks: the match may be incorrect and only bestsellers will be promoted. Therefore, the algorithm tries to collect additional knowledge about the product and the user: price filtering, interesting sizes, color, brand, etc.

This type of filtering is more accurate than other types of filtering, but it also requires more data and processing power.

Hybrid system

Recommendation systems have different types, but this does not mean that you need to use only one. There are hybrid systems in which the above algorithms can be combined and complement each other.

An example of this is TikTok. When you register a new account, you are asked to select your interests and accounts of popular bloggers. After that, TikTok will show you the most popular videos on the selected topics, and then it will take into account your interests, interaction with the content, the “Not interested” button, the similarity of your interests with other users, and so on.

When you create a new account, the content will be determined separately, regardless of the device from which you logged in. This way you can create several accounts with different algorithm settings: sports, dance, humor, and so on.

Hybrid recommendation systems are more accurate and relevant than single-type systems. They are also able to learn and adapt over time, which makes them even more effective.

How Instagram’s recommendation system works

Instagram algorithms are engaged when:

A) The user is scrolling through the feed of posts
B) Instagram Explore
C) In Reels.

Instagram also uses a recommendation system to suggest more accounts to subscribe to.

Feed

Suggested posts you see on Instagram are recommended based on your actions and the author’s actions on Instagram:

  1. Which accounts you subscribe to and which posts you interact with;
  2. The titles of the posts you interact with;
  3. How other users with similar interests interact with the posts;
  4. The date the publication was posted;
  5. Your favorites;
  6. Popularity of the publication;
  7. How often you interact with the person’s account over the past few weeks (the same principle is used to select the order in which Stories are displayed).

It is important to realize that only the accounts that the person is subscribed to will be shown in the post feed. If you want your posts to appear in the feed more often, you need to work with the interaction with the account first.

Instagram Explore

Getting into Instagram Explore recommendations is about the same thing if your video ” hits” Reels. You’ll get a lot of outside audience attention. In general, it’s not just posts that get into Explore, but Reels and even Stories themselves.

More than 200 million accounts visit the Instagram Explore page every day. That’s 50% of the platform’s users. Instagram’s algorithm ranks content based on several factors. These factors include engagement, freshness, content quality, and relevance.

Reels

Reels is undoubtedly a great tool for promoting your account. The algorithm is similar to Explore, Stories and Feed, but like the rest, Reels has its own.

Ranking Factors:

  1. User activity. Instagram keeps track of videos that people have interacted with. If a user interacts with your video, it sends signals to the algorithm to see more similar content (including videos not created by you).
  2. User history when a user interacts with your Reels. If a user continues to interact with your videos, the algorithm realizes that your content is interesting to them.
  1. Information about the video. This includes cues in the video including music, sound effects, understanding the video based on pixels and frames, etc. If you download a video from TikTok and post it to Reels along with watermarks, the algorithm will see that right away.
  2. Information about the person who posted. Similar with the feed of posts. The more a person interacts with your account, the more chance they have of seeing your new video.

How TikTok recommendation system works

We won’t be able to tell you about all the pitfalls, but we will emphasize the main principles. According to the official TikTok guide, the algorithm considers several factors before showing you a particular video on a particular topic:

They are not particularly different from other social networks and have a common principle: the algorithm adjusts to the user’s interests. Key factors:

  • Videos you like
  • Videos you share
  • Accounts you subscribe to
  • Content you create
  • Comments you post
  • Videos you add to your favorites
  • Authors you hide
  • Videos you mark as “not interesting”

Also, the TikTok algorithm reads information in some details:

  • Video description to identify any keywords that can tell the algorithm what the content is about (works especially well when searching for a specific video);
  • Hashtags to categorize content and/or inform the algorithm what the video is about;
  • Audio, such as sounds and songs, to identify trending audio content and share it with a wider audience.

Nuances of the algorithm’s work:

After publication, a video goes through pre-moderation, which can take from a few minutes to a few hours. In some cases it is automatic moderation through a robot if the account can be trusted, and in some cases it is manual. After pre-moderation, the video ends up in the feed.

The algorithm tries to understand the category of your video based on the factors above and user feedback.

The video starts to appear randomly in subscribers’ feeds, and in parallel it is shown to people from different categories. For example, a video will appear in the feed of 20 subscribers (if there are any), and 20 people from different categories (for example, 4 people who are into cars, 4 into humor, 4 into politics, etc.). If a video becomes interesting to some category, and not only likes are taken into account, but also the duration of viewing, sharers, saves, link copying, pauses, etc., TikTok increases the number of people to whom the video is displayed in that category.

YouTube Recommendation System

The main page is populated with videos that the algorithm considers interesting to the user. These can be videos from channels that the user is subscribed to or has interacted with recently, as well as channels that the user is seeing for the first time.

There are also recommendations that appear on the right side while watching a video. Here, the algorithm takes into account not only interests, but also similar videos to the one that is currently being watched.

The system takes into account the following factors to select videos that are interesting to the viewer:

  • Videos that are watched
  • Videos that are not watched
  • How much time the user spends watching the video
  • Likes and Dislikes
  • Disinterested
  • Survey results

The most important metrics taken into account by the algorithm for personalized recommendations:

  • Video effectiveness – how much a given video is liked by users with similar interests
  • User’s search and browsing history

YouTube Shorts

YouTube’s new creation and new algorithm for short videos. Let’s note right away that Shorts and long videos have different algorithms, so if your Shorts are successful, it doesn’t mean the algorithm will project success to long videos.

In the case of Trending, the same principles work as for long-form videos. The own feed focuses on interaction with videos, similar topics, and common user interests. In the case of the recommendation feed on the homepage, Shorts is geared more towards a broad audience rather than a personalized user feed.

Why you shouldn’t rely on algorithms

The essence of recommendation system algorithms is to show interest-based content to users regardless of publication time and other factors.

The scheme is simple:
Content is posted -> content appeals to a certain audience -> algorithm shows content to an audience with similar interests

This is all in theory, but it’s not that simple.

Leave a Reply

Your email address will not be published. Required fields are marked *