Hyperight

Personalization: Why and How – Ning Zhou, Microsoft

Ning will discuss which use cases have the real need of personalization and the reasons of that. Then I will go into details of one or two examples on model architecture, training strategy and performance measurement.

Key Takeaways

  • Make sure your non-personalized models work first
  • Staged training can facilitate debugging
  • Embeddings trained with user behaviours often perform better than explicit labels

Add comment