Community building recommender for group chats in LinkedIn Messaging

Adam Leon, David Golland28 Apr 2022

Problem

When a LinkedIn member builds a group chat, they must find the right people among thousands of connections by memory. People forget valued contacts, and the network mixes colleagues, alumni and event contacts in overlapping subsets. Standard pairwise recommendations do not fit because the right next person depends on who is already in the chat. The team needed suggestions that update with every added person.

Example from the article: a member with over 2,000 connections wants to reunite his university alumni, some of whom also volunteer together, and has to search for each person by hand.

Solution in brief

Build a personal bipartite graph per member: connections on one side, communities taken from the member's own profile (employers, schools, groups) on the other. At request time, walk two steps from each person already in the chat through shared communities and rank candidates by visit count. Later, replace equal counts with weights learned by logistic regression.

Questions this case answers
  • How to recommend additional members for a set being assembled when pairwise affinity cannot say which items belong together and the candidate pool is far too large to precompute for every subset
  • How to learn the relative value of different relationship types in a graph-traversal recommender when there is no logged ranking bias to rely on
  • How to recommend people to add to a group chat when the chat's purpose is a community the organizer has not named?
  • How to make recommendations update instantly as each person is added to a group?

Tools and methods

  • Personalized bipartite graph (member to community) — stores who shares which employer, school or group with the organizer
  • Two-step breadth-first traversal with visit counting — online candidate generation and ranking from the query set
  • Logistic regression on path-count features — learns how much each community type should weigh
  • Random ranker for training data — collects unbiased click labels
Continue reading — it's free

Google or a one-time email code. No password.