Overview
User Signal Service (USS) is a centralized online platform that supplies comprehensive data on user actions and behaviors on Twitter. This information encompasses both explicit signals, such as favoriting, retweeting, and replying, as well as implicit signals, including tweet clicks, video views, profile visits, and more.USS gathers signals from various underlying datasets and online services, processing them into uniform formats for consistency across the recommendation pipeline.
Signal Types
Explicit Signals
Explicit signals represent intentional user actions that directly express user preference:Favorites
User clicks the like button on tweets
Retweets
User shares tweets to their followers
Replies
User responds directly to tweets
Bookmarks
User saves tweets for later viewing
Shares
User shares tweets via external channels
Quote Tweets
User retweets with added commentary
Implicit Signals
Implicit signals capture user behavior that indicates interest without explicit engagement:Tweet Clicks
User clicks to view tweet details
Video Views
Video playback duration and completion
Profile Visits
User navigates to author profiles
Impressions
Tweets displayed in user timeline
Dwell Time
Time spent viewing content
Scrolls
Scroll patterns and depth
Data Processing Pipeline
Standardization
USS ensures consistency and accuracy by:- Collecting signals from various underlying datasets and online services
- Processing raw signals into uniform formats
- Validating signal quality and consistency
- Distributing standardized signals to downstream systems
Standardized source signals are utilized in both candidate retrieval and as machine learning features for ranking stages.
Use Cases
Candidate Retrieval
Candidate Retrieval
USS signals help identify relevant candidates from Twitter’s billion-scale corpus by understanding user preferences and historical behavior patterns.
Machine Learning Features
Machine Learning Features
Signals are transformed into features for training and inference in ranking models, providing rich context about user-tweet interactions.
Real-Time Personalization
Real-Time Personalization
Recent signals enable dynamic personalization of timelines and recommendations based on current user interests.
Model Training
Model Training
Historical signals serve as training labels for supervised learning models across the recommendation pipeline.
Signal Quality
USS maintains high signal quality through:- Deduplication: Removing duplicate events
- Validation: Ensuring signal integrity and format compliance
- Filtering: Removing spam and invalid actions
- Normalization: Consistent timestamp and identifier formats
Integration with Data Pipeline
USS integrates with other components in the data pipeline:Related Components
Unified User Actions
Source of raw user action events
Retrieval Signals
How signals are used in candidate sourcing
Aggregation Framework
Transforms signals into aggregate features