The Interest Graph


We calculate what’s interesting on the web in real-time. As new content is published across the web and shared on social networks, Gravity crawls, indexes and semantically analyzes each web page, tweet or status update. Structured and unstructured content is organized by interest so that we can determine trending topics and surface the most relevant stories about any topic on demand. Ultimately we deliver personalized recommendations to each user based on the topics they engage with most and what’s exciting right now.


Understanding the Web



Teaching Computers to Read


Natural Language Processing

  • Analyzing Content

    Artificial Intelligence and Linguistics

    We crawl and index the entire internet to figure out what people are writing about. Doing this requires computers that know how to think like humans by reading sentences and picking them apart for their grammar: nouns, verbs, adjectives, etc. Your high school English teacher would be proud of this system. We also understand things like colloquialisms: red as a lobster usually means sunburn even though those words usually map to seafood and colors.

The Math Behind Recommendations



  • True Personalization

    Context, Popularity, Virality, Recency

    Some examples are: personalization, behavioral patterns (such as collaborative filtering), contextual relevancy, popularity (by page views, uniques, social counts such as Facebook Likes, Twitter tweets/retweets). All of these can be A/B tested to determine which algorithm performs best for your particular set of content.

Improving With Feedback


Machine Learning

  • Feedback Loops

    Always Testing

    Any content that is targeted for a particular user has a feedback loop back to Gravity to tell us if that recommendation performed well or not. This informs the system so that it knows when it did a good job and when it did not. It also learns from first-time users to solve the cold start problem: how do you recommend something to someone you know nothing about?