What are the trade-offs between different loss functions for your specific recommendation use case at Criteo?
