Scalable Cross-Entropy Loss for Sequential Recommendations with Large Item Catalogs