Eitan Netzer & Oren Netzer - Real Time Machine Learning | PyData Global 2023

Core sets provide a weighted subset of data that preserves statistical properties while significantly reducing training time and compute costs

The Data Heroes Real-Time ML Framework enables:

• Automated high-frequency model retraining
• Training on multiple date ranges
• Efficient hyperparameter tuning
• Distributed processing across geographic locations

Model retraining frequency improvements:

• Monthly retraining saved 82% compute costs
• Weekly retraining improved accuracy by 22%
• Daily retraining increased accuracy by up to 20%

Core set tree structure benefits:

• Infinitely scalable and distributed
• Built once but usable multiple times
• Allows training on any subset of data
• Processing time stays consistent as data grows

Real-world case study results:

• Reduced hyperparameter tuning time from 154 hours to 5 hours
• Maintained or improved model accuracy vs full dataset training
• Enabled expansion from 144 to 864 hyperparameter combinations
• Achieved 11% average accuracy increase over 26 weeks

Implementation features:

• No data needs to leave local environment
• Works with existing ML libraries (XGBoost, LightGBM, etc.)
• Supports both classification and regression
• Automated data structure conversion