Design a Real-Time Recommendation System for Live Streams

Context: You are designing a recommender for a large live-streaming platform. Assume you have standard logs: impressions, clicks, view_events with join/leave timestamps, and stream metadata. Address the following:

1. Objective and Labels

• Define your primary objective (e.g., maximize post-click 10-minute watch probability).
• Construct training labels from view_events, including how to handle delayed outcomes and right-censoring when a stream ends before the label horizon.

2. Features

• List user, creator, stream, and context features.
• Explain cold-start handling for new creators and new users.
• Describe how you would incorporate short-term session signals and long-term embeddings.

3. Model and Loss

• Choose a model family (e.g., two-tower retrieval + re-ranker with calibrated probabilities).
• Specify the loss function(s), negative sampling strategy (in-batch + hard negatives), and how to address severe class imbalance.

4. Feedback Loops and Bias

• Mitigate position bias and popularity bias using counterfactual estimation (IPS/DR) or randomized exploration.
• Describe your exploration policy (e.g., Thompson Sampling or UCB) and safety constraints.

5. Evaluation

• Define offline metrics (AUC, PR-AUC, calibration, NDCG@k) and online metrics (watch_time/viewer, session_length, bounce_rate).
• Explain how to establish reliable offline-to-online correlation using replay evaluation and interleaving.

6. Serving and Freshness

• Provide an end-to-end latency budget (<100 ms p95) including retrieval, feature fetch, and ranking.
• Describe feature freshness (e.g., streamer going live) and how you update embeddings in near real time.