Bing Ads team User Modeling

In charge of the ranking of recommended apps for the advertising page of Xiaomi’s Appstore at Enterprise team.• Accelerated the Spark daily Profile data generation up to 3-4x by dynamic allocation, adaptive query execution and slow nodes analysis.• Built users' App preferences features in the feature engineering library. Reduced feature size significantly and brought an increase in model AUC.• Designed 20+ Apps’statistical features in the library. Features served as biases and… Show more In charge of the ranking of recommended apps for the advertising page of Xiaomi’s Appstore at Enterprise team.• Accelerated the Spark daily Profile data generation up to 3-4x by dynamic allocation, adaptive query execution and slow nodes analysis.• Built users' App preferences features in the feature engineering library. Reduced feature size significantly and brought an increase in model AUC.• Designed 20+ Apps’statistical features in the library. Features served as biases and brought 0.06% increase in performance.• Constructed user sequence features using Spark and Luigi. Investigated different sequential models including DIN. Gained 0.2% increase in model performance. Flushed features to Redis for online serving and streaming training.• Optimized model through NAS-based feature selection. Re-trained NAS-selected model from scratch and achieved an increase in model performance. Reduced input size of Tower in the model by more than 30%. Show less