Feed Based System

Problem Statement

Design a Twitter with 500 million daily active users feed system that will show the most relevant tweets for a user based on their social graph.

Metrics

• Topline is that it’s scientific as well as a business-driven decision. Overall user engagements: comments, likes, and retweets. Business-driven decision: time spent on Twitter.
• Weighted engagement + Normalized Score (divided by the total number of active users). A higher score equates to higher user engagement.

Architecture

Tweet Selection

• New Tweets + Unseen Tweets: To fetch a mix of newly generated Tweets along with a portion of unseen Tweets from the cache.
  • Edge case: User returning after a while
    • The pool of Tweets keeps on increasing so a limit needs to be imposed on the number of Tweets.
• Network Tweets + interest/popularity-based Tweets
  • Two-dimensional scheme: selecting network Tweets + potentially engaging Tweets.
  • Benefits
    • Helpful for Bootstrap problems
    • To increase the discoverability on the platform and help grow the user’s network.

Ranking

• Logistic regression
  • Pros and Cons:
    • it is fast to train
    • the major limitation of linear models is that it assumes linearly exists between the input features and prediction.
  • Approaches
    • Train a single classifier for overall engagement.
    • Train seperate models for each engagement action based on production needs (i.e. P(like), P(comments), P(tweet)).
• MART (multiple additive regression trees)
  • Tree-based models don’t require a large amount of data as they are able to generalize well quickly.
    • Train a single model for the overall engagement.
    • Train specialized predicators to predict different kinds of engagement.
    • Overall engagement + individual predictor of each engagement action -> to build one common preictor
      • P(engagement) and share its output as input into all of your predictors.
• Deep Learning
  • Seperate neural networks
  • Multi-task neural networks

• Stacking models and online learning
  • Training tree-based models and neural networks to generate features that we will utilize in a linear model (logical regression)
    • Trees generate features by using the triggering of leaf nodes (result in a boolean feature).
    • Plug-in the output of the last hidden layer as features into the logistic regression models.
    • Logistic regression: this helps the model to re-learn the weight of all tree leaves as well.
  • Main advantage: online learning
    • Using real-time online learning with logistic regression so that we can utilize sparse features to learn the interactions.

Training Data Generation

• Training data generation through online user engagement
• Balancing positive and negative samples
  • Randomly downsample
• Train test split
  • We are building models with the intent to forecast the future.

Feature Engineering

The machine learning model is required to predict user engagement on user A’s Twitter feed.

• Dense features

  • User-author features
    • User-author historical interactions
      • author_liked_posts_3months
      • author_liked_posts_count_1year
    • User-author similarity
      • common_followees
      • topic_similarity
      • tweet_content_embedding_similarity
      • social_embedding_similarity
  • Author features
    • Author’s degree of influence
      • is_verified
      • author_social_rank
      • author_num_followers
      • follower_to_following_ratio
    • Historical trend of interactions on the author’s Tweets
      • author_engagement_rate_3months
      • author_topic_engagement_rate_3months
  • User-Tweet features
    • topic_similarity
    • embedding_similarity
  • Tweet features
    • Features based on Tweet’s content
      • Tweet_length
      • Tweet_recency
      • is_image_video
      • is_URL
    • Features based on Tweet’s interaction
      • num_total_interactions
        • caveat: We need to apply a simple time decay model to weight the latest interaction more than the ones that happened some time ago.
      • Seperate features for different engagements
  • Context-based features
    • day_of_week
    • time_of_day
    • current_user_location
    • lastest_k_tag_interactions
    • approaching_holiday

• Sparse features

  • unigrams/bigrams of a Tweet
  • user_id
  • tweets_id

Diversity

• Introducing the repetition penalty as adding negative weights
• To bring the Tweet with repetition three steps in the sorted list.