Twitter has just open-sourced its algorithm, and I have carefully studied it and summarized some key points for gaining readership and posting.

Part One: Core of the Algorithm

The recommendation of tweets will be divided into three steps.

Step One: Filtering

From billions of tweets, the system will filter based on the following principles for users:

a. Followers: First, tweets from the people you follow.

b. Collaborative Filtering: For example, tweets liked by users with similar behaviors to yours, in simple terms, recommendations from people with similar tastes.

c. Second-Degree Connections: Tweets that are liked or followed by people in your second-degree connections.

d. Semantic Vector Search: The algorithm converts both "users" and "posts" into vectors in a high-dimensional space. If your interest vector is very close to the content vector of a tweet in that space, it will be selected.

This step will filter out 1,000 to 5,000 tweets. Here, it can be seen that Twitter recommends based on interests and similar behaviors, with tweets from followers only accounting for one-fourth of the total. This explains why many influencers with hundreds of thousands of followers sometimes have low tweet readership.

Step Two: Ranking

In this step, the system uses a large deep learning model like Phoenix to score each post precisely. The metrics are multidimensional, covering basically all the content we see, with a total of 15 positive indicators:

1. Favorite (Likes)
2. Reply
3. Retweet
4. Quote (Quote/Retweet with comment)
5. Photo Expand (Click to view the full image)
6. Click (Click on link/card)
7. Profile Click (Click on avatar to go to profile)
8. VQV (Video Quality View - Effective video play)
9. Share
10. Share via DM (Share via direct message)
11. Share via Copy Link (Share via copied link)
12. Dwell (Dwell time, meaning stopping on this tweet)
13. Dwell Time (Duration of dwell time, a continuous variable; the longer you stay, the higher the score)
14. Quoted Click (Clicked on the original tweet in the quoted tweet)
15. Follow Author (Followed the author after reading the tweet)

However, the weights are not disclosed on GitHub. Therefore, we cannot speculate which dimensions are more important. In other words, Twitter may dynamically adjust the weights of these dimensions.

Step Three: Filtering and Reordering

This is the final filtering of the posts ranked in the previous step. Generally, it involves compliance checks: automatically removing hate speech, pornography, violence, or deleted content. Deduplication: filtering out tweets you have already seen or those that are highly repetitive.

It is important to note that if the top-ranked tweets are from the same author, the algorithm will forcefully lower the scores of subsequent tweets from the same author, allowing you to see more diverse content.

Part Two: Summary of Key Points

1. Crossing a Threshold for Viral Posts:

The code explicitly states to prioritize in-network candidates over out-of-network candidates. This means that the coefficient for outsiders in the recommendation system is discounted, although the exact discount coefficient is not disclosed. However, the code clearly indicates that it is harder for outsiders to be recommended than for followers. This implies that to get outsiders to see your posts, you need to accumulate a certain amount of data.

2. Blue Check (Verified Badge) Weight Bonus:

In the IneligibleSubscriptionFilter, the system checks subscription status. Although the specific weighting value is hidden in the parameters, using "subscription status" as an independent feature input to the model indicates that it has a direct impact on scoring.

Moreover, Elon Musk has publicly stated that to combat bots, the visibility of tweets from unverified users in the "For You" section will be significantly suppressed by the algorithm.

3. Posting Frequency Matters:

Posting frequency can be a tricky area. There is a component in the code called AuthorDiversityScorer. Its logic is very "impartial": if you post several times in a short period, the system will consider you to be spamming, and the score of the second and subsequent posts will be directly multiplied by a decay factor.

In other words, if you post too frequently, it can be detrimental.

4. Post When Your Followers Are Active:

This is not a superstition but a definite fact. Although it is not directly in the algorithm code, it can be logically inferred as an important rule. Specifically, the Twitter recommendation algorithm includes a series of metrics such as comments and likes. If you post when your followers are asleep, the engagement data will be minimal. When your followers are active on Twitter, your post will be ranked alongside other data, and it will likely end up at the bottom.

5. Seize the "Golden Hour": Speed of Launch Determines the Ceiling:

The algorithm will first push your post to a small portion of active followers to observe their Engagement Velocity. Combined with the fourth point, it is recommended to seize the golden hour to post content at your followers' most active times. If the engagement data in the first 30-60 minutes is poor, the algorithm will determine that the post is of average quality and stop spreading it to out-of-network users.

You might consider immediately guiding core followers to engage on other social channels (like Telegram/Discord) after posting to artificially create an initial pulse signal.

6. Emphasize: Replies, Comments, User Dwell Time:

Although the weights of various metrics are not disclosed in the code, Elon Musk has publicly explained the logic of this algorithm multiple times on X, clearly stating that "the weight of replies is very high." According to the general industrial standards of recommendation systems, the weight of sparse behaviors (like replies) is certainly much greater than that of dense behaviors (like likes), otherwise, the model cannot converge.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。