Twitter makes the keys of its algorithm public

Which division is currently investigating a leak? source code 100% of the social network accessed the portal in question published by a user identified as ‘FreeSpeechEnthusiast’ without permission.

GitHub has removed this content from its platform after receiving an email. judicial notice Due to copyright infringement posted by Twitter, which then requests that both the person who leaked it and those who downloaded it be known.

This is just days after Twitter owner Elon Musk announced it at the end of March.or make the platform’s algorithm publicSomething that finally officially happened on Friday, March 31.

The platform has shared this content for the purpose of “giving”. the first step in a new era of transparency” and to exclude “any code that could compromise the user’s security and privacy,” as he explains in a statement he shared on his blog.

GitHub now has two new repositories – ‘Main’ and ‘ML’; determines which “tweets” the social network offers To users in the “For you” section.

Twitter also stated that it has shared more information about the recommendation algorithm on its blog. How do you route suggestions and how do you filter suggestions? person considering related posts to recommend them.

However, despite making this part of the source code available to users, decided to skip the section devoted to ad suggestions.

recommendation algorithm

Platform, provides information to users and select a series of ‘tweets’ to add to the ‘For you’ section of each.

Twitter approves recommendation system “Consists of many interconnected services and jobs” revealed how this information is filtered through a process consisting of three stages.

First, it takes the “best tweets” from different referral sources – in a process called “lead sourcing” – and second, it ranks each one using a machine learning algorithm. (‘machine learning’).

Finally, it applies heuristics to filter tweets, posts you’ve seen before, and Content tagged as safe or unsuitable for business (NSFW).

Name of the service responsible for creating the ‘for you’ section home mixer (a kind of homepage mixer) and the Product Mixer program that “acts as the backbone” of the “software” that makes it easy to create content sources and connects the lead posts and other rating that will be part of this section. functions.

The first phase of the process considers a set of resources it extracts. top 1,500 tweets from a pool of hundreds of millions of these resources. To do this, it uses accounts that are both followed and unfollowed by users.

“Today, the ‘For You’ chronology consists of: 50 percent of tweets are on the network, 50 percent are off the network On average, this can vary from one user to another,” the social network described.

Regarding the resources that users follow, Twitter, “top candidate” He states that his aim is to show the most up-to-date and relevant ‘tweets’ of followed users.

That’s when it uses a tool called real chart, a machine learning model that predicts the probability of interaction between two users. The more compatibility there is, the more ‘tweets’ will be included.

The platform announced that it was working on this part recently as it stopped using it. Fanour Servicea solution previously used to boost posts from a cache for each user.

Unfollowed users

The social network has also explained how it integrates the feeds of unfollowed users in their ‘For you’ recommendations and furthered it. take two positions around it.

First, it takes into account the so-called social Graphic, it tries to answer which tweets the users of the people they follow interact with and who likes the posts they make.

After these questions are answered, the candidate creates “tweets” based on these answers and categorizes them using a logistic regression model. To know his journey on the platform, he developed a graphics processing engine named after him. GraphicJet.

Another approach that Twitter is committed to suggesting posts from unfollowed accounts is the so-called ‘placement spaces’‘Which tweets and users are similar to my interests?’ aiming to answer the question.

These burials user interests and examines the content of the ‘tweets’ and then calculates the similarity between two users or two random ‘tweets’ integrated into this field.

The next step is to categorize the proposed tweets, a point in the process where approximately 1,500 candidates are submitted. “may be relevant” and each candidate receives a score that directly estimates the importance of the publication.

This classification is obtained with a neural network of approximately 48 million parameters and is continuously trained. optimize positive interaction on the platform. That’s when the system generates ten tags to give each ‘tweet’ a certain score. Each represents the probability of interacting with those posts.

Apply after this classification a set of filters this allows the platform to make more precise recommendations to deliver various results. Currently, Twitter is removing suggestions from blocked accounts, reducing the number of consecutive ‘tweets’ from a single account, among other actions.

At the endpoint and with suggestion posts already selected, Twitter activates Home Mixer and sends the suggestion to each device. At this point in the process, the classification system combine ‘tweets’ with other content, like ads or follow suggestions to other accounts.

After publishing this part of the Twitter algorithm for recommendations, Musk confirmed in his profile on the social network: They will unlock anything that contributes to the viewing of a tweet in the “coming weeks”.

The company also confirmed its intention to expand the referral system with new opportunities that they are already working on. new features in real timeembedded and user representations.