Learn→Recommendation Systems→Real-time Personalization (Session-based, Contextual Bandits)→1 of 6

Recommendation Systems • Real-time Personalization (Session-based, Contextual Bandits)Easy⏱️ ~3 min

What is Real-time Personalization?

Definition
Real-time Personalization adapts recommendations within seconds based on what a user does during their current session, rather than waiting for batch model updates that take hours or days.
THE CORE PROBLEM
Traditional recommendation systems train models on historical data and serve predictions from static user profiles. When a user starts searching for running shoes, the system keeps recommending office supplies based on past purchases. The next batch training cycle might be 6 to 24 hours away. By then, the user has left. Real-time personalization captures and acts on intent within the same session.
TWO MAIN APPROACHES
Session-based models: Track user actions within a visit and predict what they want next. A user viewing hiking boots, then socks, is likely interested in hiking poles. The model learns these sequential patterns. Contextual bandits: Balance showing items the system knows work well (exploitation) versus trying new items to learn preferences (exploration). Both adapt within seconds rather than waiting for batch retraining.
WHY IT MATTERS
Users often arrive with intent that differs from their historical profile. A parent buying toys looks nothing like their typical purchases. Capturing session intent increases conversion by 15 to 30% compared to relying only on history. The tradeoff is infrastructure complexity: real-time systems need streaming feature pipelines and sub-100ms response times.
💡 Key Insight: Real-time personalization is not about faster models. It is about incorporating signals that only exist during the current session.

💡 Key Takeaways

✓Adapts recommendations within seconds based on current session, not batch updates taking 6-24 hours

✓Session-based models predict from action sequences; contextual bandits balance exploitation and exploration

✓Captures intent that differs from historical profile, like a parent buying toys

✓Increases conversion by 15-30% compared to history-only recommendations

✓Requires streaming feature pipelines and sub-100ms response times

📌 Interview Tips

1Explain the lag problem: user searches running shoes but sees office supply recommendations for hours

2Describe session sequence: hiking boots → socks → poles shows clear intent progression

3Discuss the 15-30% conversion lift and when it justifies the infrastructure investment

← Back to Real-time Personalization (Session-based, Contextual Bandits) Overview