Personalization without user identity

This post introduces Proximity Features, a novel ML feature set designed to address the cold-start problem in personalization for users without individual history or identifiers. It details the core hypothesis of using aggregate activity patterns within geographic proximity buckets. The technical contributions include: 1. Definition of a 'proximity key' as an aggregation key for grouping ~1,000 users based on quantized latitude/longitude tiles and IP hash buckets. 2. Development of a two-phase adaptive clustering algorithm for stable, zoom-adaptive bucket creation that handles varying user densities. 3. Computation of three categories of features: short-term engagement, long-term booking patterns, and aggregate bucket metadata. 4. A privacy-by-design approach, excluding unconsented users and using coarse geo-IP aggregation. 5. Operationalization of these features for real-time inference via a distributed key-value store, with a soft dependency for reliability. The post also presents production A/B experiment results demonstrating measurable lifts in personalization for cold-start traffic.