際際滷

際際滷Share a Scribd company logo

Suggests

Download as PPTX, PDF
T
TheDataNation

Trulia collects large amounts of data from its 31 million users and other sources to power personalization features on its real estate search platform. It developed an early prototype for personalized suggestions based on Netflix Prize models but found a better approach was to model user click behavior for real-time recommendations. The next version of Trulia Suggests will more closely integrate with the home buying process, use additional models beyond preference prediction, and provide a more visual and entertaining experience.

1 of 21
Downloaded 15 times
Trulia Suggests April 4, 2013
Trulia
Big Data @ Trulia 31M users Property data Images Public record data Agent data Local data (crimes, transit, etc) Q&A data
Trulia
Trulia
Trulia
Trulia
Natural Fit for Personalization One (long) search Sometimes very long (see: San Francisco market) Often multiple sessions Lots to interact with on Trulia
Largest Companies
Limited Personalization
Prototype Initially considered work the work surrounding the Netflix Prize and KDD Cups
Prototype Quickly recognized - real-time higher priority than accuracy - requiresbetter model of user click behavior
Prototyping Unconstrained interface development
Hide, Like, or Follow
Pick a house user community
Pick a house user community
Pick a house user community
Suggestions Visual Interface
Suggestions Assembling Lists
Trulia Suggests v2.0 More integration into the natural home buying process More models, not just preference prediction More visual More entertaining
Trulia Were hiring! datascience@trulia.com

More Related Content

Suggests

Editor's Notes

  • #9: Everything logged
  • #12: We were familiar with this work, which focuses on preference prediction, and without time constraints.
  • #13: Users are easily distracted
  • #14: The interface decisions are at least as important as the algorithm decisions. Particularly early on. We didnt want to be biased by our current FE (FE tech has developed rapidly) so we build a separate prototype using the latest tech and then found ways to carry elements back.
  • #15: Hide because people often do repeat searches, like as the positive preference, and follow as a stronger like in which updates are sent
  • #16: We originally planned to bootstrap the system using only behavior. But users are easily distracted. While we continue to build better models of user real estate search behavior, we decided just to ask users.
  • #19: Recommendations themselves are displayed in a visual interface. Cards instead of lists. Inline photos. Bits of animation.
  • #20: Lists are central to home buying. Already on IPad. Coming soon to rentals, mobile.