KOEN
Read Original
#AI

Personalization without user identity

Personalization without user identity
01

Summary

Personalization Without Cookies: How Airbnb Uses Proximity Signals to Beat the Cold-Start Problem

A clever privacy-first ML architecture that group-aggregates anonymous traffic behavior instead of individual tracking

Airbnb introduces 'Proximity Features', a novel framework that personalizes services for anonymous or logged-out users without relying on individual history or tracking cookies. By mapping coarse IP locations into adaptive buckets of roughly 1,000 local users, Airbnb feeds group-level preferences to machine learning models, successfully overcoming the privacy era's toughest personalization constraints.

  • 01A highly creative alternative to cookie tracking that implements group-level crowd signals to power machine learning models.
  • 02An elegant two-phase Adaptive Clustering algorithm that automatically zooms in for metropolitan areas and zooms out for rural terrains.
  • 03Compliance-by-design pipelines integrated with consent and data governance controls to adhere with GDPR.
  • 04High availability ensured through a soft-dependency architecture, preventing core user path blockages during feature store latency spikes.
  • 05Empirically validated lifts in user conversion and recommender diversity among never-booked and dormant populations.

RECOMMENDATION

Highly recommended for data engineers and ML system architects struggling with the death of third-party cookies and tightening GDPR guidelines on user-level tracking.

The Problem

Recommender and search ranking systems face a severe cold-start problem with guest users or logged-out audiences because no individual tracking history is available. Additionally, the deprecation of third-party cookies and modern privacy compliance frameworks like GDPR have limited the feasibility of using persistent tracking identifiers.

The Solution

Airbnb developed 'Proximity Features' by grouping geographically adjacent users into adaptive buckets of ~1,000 via a two-phase 'Adaptive Clustering' algorithm. This constructs localized aggregation keys ('Proximity Keys') using coarse IP-derived locations to serve real-time localized recommendation features computed on daily batch pipelines.

The Result

Production A/B testing on marketing landing pages and Homepage AutoSuggest yielded strong engagement and search-term diversity lifts, successfully shifting the fallback suggestions from generic global lists to highly localized alternatives for new and dormant users.

Trade-off

Retrieving proximity features from the key-value store operates as a soft dependency, meaning latency timeouts will silently fall back to default non-personalized outputs, sacrificing personalizing guarantee for service availability. It also compromises on specific individual preferences by relying strictly on collective crowd signals.

03

Key Concepts

Concept · 01

Proximity Key

A compact geographic group identifier representing approximately 1,000 nearby users, calculated from quantized lat/long coordinates.

  • Replaces user_id in machine learning model inferences when no logged-in user context is available.
  • Balances granular tracking constraints by encoding coarser, group-level patterns rather than an individual address.
Concept · 02

Adaptive Clustering

A spatial partitioning algorithm designed to cluster global users into stable geographic groupings of ~1,000, regardless of local population density.

  • Subdivides dense urban locations with IP hashing and merges sparse rural locations into larger geographical zones.
  • Generates robust, production-stable clusters that can persist with daily incremental refreshes.
Concept · 03

Soft Dependency

A software design pattern where the failure or slow response of optional data retrieval does not halt or fail the main execution path.

  • Implemented in the real-time serving layer during proximity feature lookups in the distributed Key-Value store.
  • Prevents checkout and navigation flows from crashing by serving generic, static global recommendation cards upon timeout.