• Twitter published some version of their "For You" homepage algorithm last week
• No model weights or data
• Includes names of features, labels, aggregation of labels into scores, and (hard to interpret) ranking logic

Please note we have force-pushed a new initial commit in order to remove some publicly-available Twitter user information. Note that this process may be required in the future.

Open Sourcing "The Algorithm"

• Notes from blog post for overall structure
• Filling in details based on code in repository

Methods

Tweet Sources

• Tweets from users I follow (around 50% of feed)
• Tweets from interaction network (around 15% of feed)
• Tweets with similar embedding to me (around 35% of feed)
• Always filtered to ensure no more than two degrees of separation

Tweet Sources

• Tweets from users I follow (around 50% of timeline)
• Tweets from interaction network (around 15% of timeline)
  • From real time graph between users and Tweets
  • Tweets that people I follow engaged with
  • Tweets engaged with by people who engage with similar Tweets as me
• Tweets with similar embedding to me (around 35% of timeline)
  • e.g. SimClusters which uses a matrix factorization algorithm on a large binary matrix of users/tweets (based on likes?)
  • (not sure how use) Twihn learns embeddings based on graph data (follow, likes), have released anonymized and smaller version on HuggingFace

SimClusters

• paper
• Sparse and interpretable
• Latent factors are communities, determined by a small number of major influencers
• Scale: \(10^9\) users, \(10^{11}\) edges between them, \(10^8\) new tweets per day, \(10^9\) user engagements per day.
• Representation space is \(10^5\): communities of around 100 users
• Step 1: "known for" one hot
• Step 2: "interested in" based on 1
• Step 3: producer embedding from 2
• Online tweet embedding is sum of interested in embeddings
•  

Lightweight model

• Logistic regression model
• Last trained/updated several years ago
  • still has the feature flag "is vine"
• Largely unpersonalized, most features pertain to tweet rather than to me
• Predicts engagement: is_clicked, is_favorited, is_open_linked, is_photo_expanded, is_profile_clicked, is_replied, is_retweeted, is_video_playback_50
• Score is weighted sum of predictions
• Top ~1500 scoring Tweets go to next step

First stage

• Logistic regression predicts relevance of tweets from candidate pool (slightly different features/models for in and out of network tweets)
  • last updated/trained several years ago
• Real Graph predicts likelihood of engagement between two users
  • features about edge (timeseries interactions) and each user (activity and reputation)
  • unclear if this is used in the light ranker, or as a candidate generation step
• Results in ~1500 tweets

Logistic regression

• labels: is_clicked, is_favorited, is_open_linked, is_photo_expanded, is_profile_clicked, is_replied, is_retweeted, is_video_playback_50
• highlights from features
  • weighted and decayed engagement (likes/replies)
  • text analysis (length, readability, offensiveness - static; toxicity - realtime)
  • encoded_tweet_features.has_vine_flag RIP
  • properties of tweet author, like "tweepcred" (pagerank for users as producers)
  • doesn't seem very personalized to the consumer beyond language/time

Heavy Ranker

• Deep network (MaskNet) ~48M parameter
• Thousands of features:
  • historical engagement between various relevant entities aggregated over various timescales
  • static properties of tweet (text, media)
  • embeddings of user, tweet (generated by link prediction)
• Output: engagement
  probabilities

Second stage

• Rank candidate tweets on basis of score (relevance)
• ~48M parameter deep network outputs one of ten labels representing the probability of an engagement
  • "optimize for positive engagement (e.g. Likes, Retweets, and Replies)"
  • thousands of features

Heavy Ranker Inputs

• Features aggregate over short and long timescales:
  • author features, tracking how their tweets are engaged with on different timescales
  • author-topic engagement on 50 day scale
  • user-list engagement on short and long scale
  • engagement by user by tweet property over different timescales
  • engagements between user and authors over different timescales
  • interactions between user and other engagers in the same tweets on long scale
  • user-topic engagement for inferred topics 50day
  • user-media interaction 50 day
  • interactions between user and other tweet mentioning the users mentioned in the tweet in question? 50 day
  • aggregate user engagements over same day of week / hour of day 50d
  • user-topic engagements (by tweet property) on 50 day
  • aggregate engagement on the tweet on multiple timescales

Heavy Ranker Inputs

• Features that aren't aggregates:
  • which two-hop relationship(s) describes tweet author and user
  • network properties about author and user from realgraph
  • also about users mentioned in the tweet, in-network engagers, and upstream authors
  • tweet's features (engagements, "has media", "has question"), users device
  • features about upstream tweet, if its a reply
  • same features as light ranker
  • more realtime user-author interaction features (less of them?)

  • similarity of tweet to user's recent engaged tweets "sim_clusters"

  • other misc features about time, request context, author health

  • Thwin embeddings: directional follow embeddings, engagement embedding, each 200 dimensional

TwHIN Embeddings

• Learned by link prediction with log likelihood of edge \(r\) between source \(s\) and target \(t\) parametrized as $$f(s,r,t) = (\theta_s+\theta_r)^\top \theta_t$$
• Paper

Heavy Ranker Model

• MaskNet with parameters
•  

categorical input fields are embedded and numerical ones are compressed to \(k\) dimensions

• \([x_1,x_2,\dots,x_f] \to [e_1,\dots, e_f]\)

Heavy Ranker Outputs

• Outputs binary predictions, which are weighted into a score and then ranked (later reranked)
  • favorite: 0.5
  • max(click and engage, click and linger 2min): 11
  • negative feedback (block, mute, "show less often"): -74
  • click to profile and engage: 12
  • reply: 27
  • reply with something the author engages with: 75
  • report: -369
  • retweet: 1
  • watch half a video: 0.005

Heuristics

• Remove Tweets based on my explicit filters, negative feedback
• Prevent too many consecutive Tweets from same user
• Balance in- and out- of network tweets
• If Tweet is a Reply, thread together with original Tweet
• Include ads, Follow recommendations, and onboarding prompts

Heuristics

• Remove Tweets based on my explicit filters
• Prevent too many consecutive Tweets from same user
• Balance in- and out- of network tweets
  • is the mix of out- because embedding is more effective?
• Lower score of tweets which have received negative feedback from me
• Exclude Tweets without 2nd degree connection (follow or engagement by someone you follow)
• If Tweet is a Reply, thread together with original Tweet
• Update Tweets which have been edited
• Include ads, Follow recommendations, and onboarding prompts

Heuristics

• I think controlled by filter and score
• Also some stuff in visibility models
  • downranking rules
  • "ForEmergencyUseOnly.scala"
  • "FreedomOfSpeechNotReach.scala"

Finally

• Show the Tweets in order, mixed with ads, Follow recommendations, and onboarding prompts
• The pipeline above runs approximately 5 billion times per day and completes in under 1.5 seconds on average. A single pipeline execution requires 220 seconds of CPU time, nearly 150x the latency you perceive on the app.