1. Architecture of the recommender system
- Item feature extraction
- photos, videos and news
- User feature/relation extraction
- Online/Offline 召回
- Trade-off between accuracy and latency
- 粗排，精排
- Graph embedding for knowledge graph? social graph?

2. Metrics
- Pupular/practical offline metrics (e.g. AUC, recall and precision, and the weighted sum of them)
- Online metrics (用户留存, A/B testing results?)
- Practical procedure (offline metrics as experiments, online test for final decision?)

3. Training
- Training on unbalanced data (正例很少，如何扩充)
- Sampling for speedup and unbalance data

4. Uncerntainty estimation
- help on cold-start