Design place-of-interest ML system
Overview
This question evaluates proficiency in designing large-scale ML recommendation systems, covering personalization, freshness, low-latency serving, data and feature engineering, training and serving pipelines, exploration/exploitation strategies, evaluation metrics, and trust-and-safety considerations.
Design a POI (Places of Interest) Recommendation System
Context
Design a global POI recommender for a mobile maps/feed product that suggests nearby places (e.g., restaurants, attractions) across surfaces such as a home feed, map viewport, and search results. The system must support personalization, freshness, and high scale while meeting strict latency targets.
Specify
(a) Product goals and key requirements:
- Personalization: individualized to user tastes, intents, and context
- Freshness: reflect open/closed status, trending, events, new places
- Latency: responsive on mobile; include p50/p95 budgets
- Scale: global POIs, high QPS, multi-region deployment
(b) Data and features:
- Data sources: map metadata, reviews, check-ins, GPS pings, events
- Labeling strategy: define positives/negatives, counterfactual logging, debias for position/exposure
- Feature sets: user, POI, context, interaction, geographic features
(c) Architecture:
- Two-stage retrieval: candidate generation (embeddings/ANN) and ranking (GBDT or deep)
- Re-ranker for diversity and novelty
(d) Training and serving:
- Batch + streaming updates, feature store, backfills
- Online serving: feature retrieval, caching, latency budgets, fallbacks
(e) Exploration/exploitation:
- Strategy (e.g., bandits, epsilon-greedy) for cold start and long-term learning
(f) Evaluation plan:
- Offline metrics (AUC, NDCG, coverage)
- Online A/B metrics (CTR, save/visit rate, dwell)
(g) Trust & safety:
- Privacy, abuse/spam prevention, geo-specific fairness considerations
Solution
Login required