Summary Twitter's Recommendation Algorithm blog.twitter.com
1,792 words - html page - View html page
One Line
Twitter's recommendation algorithm selects top tweets using a logistic regression model based on social graph to deliver a diverse and relevant feed to users, and the company is expanding its features and transparency while building the town square of the future.
Key Points
- Twitter serves over 150 billion Tweets to users daily and is expanding their recommendation systems with real-time features, embeddings, and user representations.
- The recommendation algorithm employs several heuristics, filters, and product features to deliver a balanced and diverse feed of relevant tweets to its users.
- The algorithm uses numerical representations of user interests and tweet content to generate accurate embeddings and a graph processing engine to maintain a real-time interaction graph between users and tweets to execute traversals.
Summaries
202 word summary
Twitter's recommendation algorithm selects top tweets from the 500 million daily to deliver the best of what's happening in the world. It generates candidate tweets and ranks them using a logistic regression model based on a social graph to estimate relevance. The algorithm uses in-network and out-of-network sources to retrieve recent and relevant tweets for a user, and a home mixer service connects different candidate sources to construct and serve the "For You" timeline. The algorithm aims to deliver a balanced and diverse feed of relevant tweets to users by employing several heuristics, filters, and product features. It uses a neural network to rank tweets on a user's timeline, embeds tweets into communities by looking at their current popularity, and employs a graph processing engine to maintain a real-time interaction graph between users and tweets. Twitter is expanding their recommendation systems with real-time features, embeddings, and user representations, and aims to provide greater transparency to users. The home mixer blends together tweets with other non-tweet content like ads, follow recommendations, and onboarding prompts before they are returned to the user's device for display. Twitter is building the town square of the future and is looking for individuals interested in joining their team.
463 word summary
Twitter serves over 150 billion Tweets to users daily and is working on expanding their recommendation systems with real-time features, embeddings, and user representations. They aim to provide greater transparency to users by allowing them to understand their algorithm in greater detail and providing visibility into why Tweets appear on their timeline. They have also released the code powering their recommendations for users to view. Home Mixer blends together Tweets with other non-Tweet content like Ads, Follow Recommendations, and Onboarding prompts before they are returned to the user's device for display. Twitter is building the town square of the future and is looking for individuals interested in joining their team. Twitter's recommendation algorithm aims to deliver a balanced and diverse feed of relevant tweets to its users. The algorithm employs several heuristics, filters, and product features to achieve this goal. The ranking stage is accomplished with a neural network that takes into account thousands of features and outputs ten labels to give each tweet a score. The scoring mechanism directly predicts the relevance of each candidate tweet and is the primary signal for ranking tweets on a user's timeline. Additionally, the algorithm embeds tweets into communities by looking at their current popularity in each community. The more users from a community like a tweet, the more that tweet will be associated with that community. To generate accurate embeddings, the algorithm uses numerical representations of user interests and tweet content. Finally, the algorithm employs a graph processing engine to maintain a real-time interaction graph between users and tweets to execute traversals. Twitter uses a recommendation algorithm to generate candidate tweets and rank them using a logistic regression model. They use a social graph to estimate what users would find relevant by analyzing the engagements of people they follow or those with similar interests. The In-Network source is the largest candidate source and delivers the most relevant, recent tweets from users you follow. Twitter also uses Out-of-Network sources to find relevant tweets outside of a user's network. Candidate sources are used to retrieve recent and relevant tweets for a user. The Home Mixer service connects different candidate sources, scoring functions, heuristics, and filters to construct and serve the For You timeline. The recommendation pipeline is made up of three main stages that consume features to answer questions accurately and deliver more relevant recommendations. Twitter's recommendation algorithm is based on core models and features that extract latent information from Tweet, user, and engagement data to determine the probability of user interaction in the future. The system is composed of interconnected services and jobs, with a focus on the home timelines "For You" feed. The algorithm selects top Tweets from the roughly 500 million posted daily to deliver users the best of what's happening in the world.