Understanding a diverse user base with frequency segmentation at scale

Cuong Duong, Naomi Stuart16 Mar 2023

Problem

Canva's users range from marketers who post several times a week to casual users who return once every few months. The company wanted to move users toward more frequent use and needed a frequency bucket per user to size and track that. The Buy-Till-You-Die model BG-NBD (Beta-Geometric/Negative-Binomial) reacts slowly: users who start returning often are still predicted as rare, because old long gaps dominate their average. Inactive users also have to be told apart from infrequent ones.

Example from the article: two real users who began returning several times a week, with gaps of at most two weeks, were predicted by BG-NBD to return only monthly or quarterly.

Solution in brief

Estimate frequency as the average of the last five gaps between active days. Divide days since last active by it to get a recency ratio, and map the ratio to churn probability with a power curve. Combine the frequency bucket, alive or potential-churn status and a first-month flag into readable segments, computed in SQL.

Questions this case answers
  • How to assign each user a usage-frequency bucket that reacts quickly to behavior changes and separates inactive users from infrequent ones
  • How to define a churn signal for users with very different usage cadences without a fixed inactivity cut-off
  • How to segment users by how often they return when usage ranges from several times a week to once a year?
  • How to tell a user who has become less frequent from a user who has churned?

Tools and methods

  • Sliding-window frequency — mean of the last five inter-activity gaps.
  • Recency ratio — flags users who dropped out of their usual rhythm.
  • Power curve — maps recency ratio to churn probability.
  • BG-NBD — baseline that was tested and rejected.
  • dbt — runs the SQL pipeline in Snowflake.
Continue reading — it's free

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