Nextdoor Notifications: Using ML to Keep Neighbors Informed
Article Summary
Karthik Jayasurya from Nextdoor reveals how they transformed notifications from random spam into personalized updates that neighbors actually want. The result? A 40% jump in engagement without sending more messages.
Nextdoor's ML team rebuilt their entire notification system to solve a critical problem: how do you keep millions of neighbors informed without overwhelming them? They replaced simple heuristics with XGBoost models and PID controllers to predict relevance and control volume at scale.
Key Takeaways
- XGBoost model increased email clicks 40% and push taps 23% at same volume
- PID controllers dynamically adjust thresholds to hit personalized weekly budgets
- System uses embeddings, real-time metrics, and proximity features for scoring
- Dormant users now get more notifications while active users get fewer
- Overall platform saw 8% increase in daily active neighbors
By combining ML relevance scoring with control theory for volume management, Nextdoor delivered dramatically better engagement while respecting user preferences and reducing notification fatigue.
About This Article
Nextdoor's notification system picked random eligible neighbors without considering what they actually cared about. It didn't explain why a post might matter to someone, so the company couldn't really understand which posts would resonate with which users.
Karthik Jayasurya's team built an XGBoost model that looked at content features, how recipients engaged with posts, embeddings between authors and recipients, and real-time interaction metrics. They then added PID control theory on top to adjust personalized notification thresholds based on each user's weekly budget.
Notifications became more fairly distributed. Neighbors who rarely engaged saw 23% more push notifications they actually clicked on, while active users unsubscribed less often. The system kept notification volume consistent across all channels.
