Recommendation System Deep Dive (Ranking, ANN, Feature Store, Real-time Personalization)

#Glossary (Thuật ngữ nâng cao)

• Candidate: tập item ban đầu
• Ranking: sắp xếp kết quả
• Embedding: vector biểu diễn user/item
• ANN: Approximate Nearest Neighbor (tìm gần đúng)
• Feature store: nơi lưu feature
• Online learning: update realtime
• Offline learning: training batch

#1. Bản chất hệ thống

Recommendation = Multi-stage decision system

#2. Multi-stage Architecture (Production)

Candidate → Pre-ranking → Ranking → Re-ranking

#Giải thích

• Candidate: lấy ~1000 item
• Pre-ranking: giảm còn ~100
• Ranking: model chính
• Re-ranking: rule business

#Insight

Multi-stage giúp giảm latency + tăng accuracy

#3. Candidate Generation (Deep)

#3.1 Embedding

• user/item → vector

#3.2 ANN Search

• tìm item gần nhất trong vector space

#Trade-off

#4. Ranking System (Deep)

#Formula

score = model(user, item, context)

#Feature

• user behavior
• item info
• context (time, device)

#Business Logic

• boost ads
• boost new content

#5. Feature Store (Deep Dive)

#Online Feature

• dùng cho serving (low latency)

#Offline Feature

• dùng training

#Problem

• mismatch giữa training và serving

#Fix

• shared feature store

#6. Real-time Personalization

#Flow

User action → Event → Update feature → Update ranking

#Example

• user click → update interest ngay

#7. Online vs Offline Learning

#Hybrid

• train offline + update online

#8. A/B Testing System

#Mục tiêu

• test model

#Flow

• user split group
• compare CTR

#9. Metrics System

• CTR (click)
• retention
• watch time
• revenue

#10. Failure Matrix (Cực sâu)

#11. Real Incident

#Case 1: CTR drop

• deploy model mới

→ CTR giảm mạnh

#Fix

• rollback model

#Case 2: Feature lỗi

• data sai → recommend sai

#12. Scaling Strategy

#13. Cost

• compute model rất tốn

#Fix

• multi-stage
• cache result

#14. Observability

• CTR
• latency
• error rate

#15. Code Example

$items = RecommendService::get($userId);

#16. Checklist Production

• multi-stage ranking
• embedding + ANN
• feature store
• A/B testing
• monitoring

#17. Final Architect Insight

Recommendation system = decision engine của product

Trade-off:

• accuracy
• latency
• cost