Apr 03, 2023 14:05:00

Posts that are easy to be posted on the ``recommended'' timeline found in the source code of Twitter's published algorithm Summary of posts that are difficult to be posted

The post selection algorithm for Twitter's 'Recommended' timeline was released on March 31, 2023, revealing what kind of posts are likely to be posted on the 'Recommended' timeline.

Twitter's Recommendation Algorithm

https://blog.twitter.com/engineering/en_us/topics/open-source/2023/twitter-recommendation-algorithm

Twitter open-source recommendation algorithm code
https://www.bleepingcomputer.com/news/technology/twitter-open-sources-recommendation-algorithm-code/

Twitter's newly-released algorithm shows it specifically tracks how Elon Musk's tweets are doing |
https://mashable.com/article/twitter-releases-algorithm-showing-it-tracks-elon-musk-tweets

How the Twitter Algorithm works in 2023
https://steventey.com/blog/twitter-algorithm

Twitter's 'recommended' timeline also displays tweets from accounts that users do not follow, but it is unclear what algorithm selects tweets, and it is unclear whether Twitter manipulates the number of views for some users. There was a suspicion that In response, Elon Musk, CEO of Twitter, declared, ``Ensuring the transparency of the code and gaining the trust of users,'' and on March 31, 2023, the ``recommended'' timeline post selection algorithm has been published.

Twitter open-sources and publishes the post selection algorithm of the 'recommended' timeline, anyone can download it-GIGAZINE

According to Twitter, the tweets posted on the 'recommended' timeline are selected by the following algorithm, and 1,500 tweets are extracted from hundreds of millions of tweets for each request. It is also stated that the extraction is adjusted so that posts by users who are following and posts by users who are not following are half and half.

1. Collecting the best tweets that can be “recommended” from various aspects
2. Use machine learning to rank collected tweets
3. Filter based on read and blocked status

The specific factors used for ranking can be found in the code on Github , which reveals the following:

◆ The score exists for the user itself
If the user's score is high, it will be easier to be displayed in the follow recommendation column, and the posted content will be easier to be highlighted. A user's score is calculated based on the following factors:

・The number and quality of interactions with other users
・Account creation date, number of followers, device usage
・Follow/follower ratio

◆ Twitter Blue users are boosted
Users who pay for Twitter Blue will be more likely to be recommended, and the score multiplier will be 4x for followers and 2x for other users .

◆ Reply to the reply will greatly increase the score
Every time a user takes action on a tweet, the probability of that tweet being recommended increases . The weighting for each specific action is as follows.

Likes: 0.5
Retweets: 1
Click a Tweet to Like or Retweet: 11
Stay 2 minutes on a tweet: 11
Like or reply to the tweet and visit the poster's profile:12
Reply: 27
Reply/Like/Retweet by the poster: 75

Select 'Not interested in this tweet': -74
Block/mute poster: -74
Spam Reported: -369

◆ When blocked or muted, the user's score drops sharply

A user's score is reduced when they are blocked/muted, reported for abuse, spammed, or unfollowed. In addition, in the case of unfollowing, the impact is smaller than in other cases, and the impact disappears after 90 days.

◆ There is a government-only command
The government seems to be able to display election candidates in the follow recommendation column and suppress the display of fake news .

◆ Images and videos are advantageous
Tweets with embedded images or videos get double the score .

◆ General users' tweets can only be evaluated up to 3 times
For regular users, the ranking algorithm only calculates up to 3 tweets. On the other hand, if the user's score is high, the restriction will be lifted and all tweets will be evaluated.

◆ Tweets have a lifespan
The tweet score will decrease at a pace that halves in 6 hours .

◆External links are dangerous for new accounts
If a tweet posted by a user who does not have a sufficient user score has external links, tweets without likes, retweets, and replies will be treated as spam .

◆ Certain categories are ranked low
Misinformation, violence/hate, and posts about Ukraine rank low.

A misspelling can be fatal
If you post in English, if you make a mistake in spelling, there is a risk that the language cannot be determined and your score will be 1/100 .

Elon Musk has said that he will update the algorithm released this time according to user feedback, and on April 1st, the day after the release , he will immediately distinguish whether the tweeter is Elon Mask or not. code has been removed .