HyperGBM Versions Save

This version brings the following new features:

• Full pipeline GPU acceleration
  • Data adaption
  • Data cleaning
  • Feature selection
  • Data drift detection
  • Feature selection(2nd stage)
  • Pseudo labeling(2nd stage)
  • Optimization
    • Data preprocessing
    • Model fitting
  • Model ensemble
  • Metrics
• Model training
  • Add TargetEncoder for categories
  • Set estimator eval_metric based on experiment reward_metric
• Advanced Features
  • Data adaption in experiment
• Experiment Visualization
  • Experiment configurations
  • Dataset information
  • Processing information
• Multijob management
  • Series and parallel jobs scheduling
  • Local and remote jobs execution
• Export experiment report

Upgrade requirements

This release add following new features:

Feature engineering

• Feature generation
• Feature selection

Data clean

• Special empty value handing
• Correct data type
• Id-ness features cleanup
• Duplicate features cleanup
• Empty label rows cleanup
• Illegal values replacement
• Constant features cleanup
• Collinearity features cleanup

Data set split

• Adversarial validation

Modeling algorithms

• XGBoost
• Catboost
• LightGBM
• HistGridientBoosting

Training

• Task inference
• Command-line tools

Evaluation strategies:

• Cross-validation
• Train-Validation-Holdout

Search strategies

• Monte Carlo Tree Search
• Evolution
• Random search

Imbalance data

• Class Weight
• Under-Samping
  • Near miss
  • Tomeks links
  • Random
• Over-Samping
  • SMOTE
  • ADASYN
  • Random

Early stopping strategies

• max_no_improvement_trials
• time_limit
• expected_reward

Advance features:

• Two stage search
  • Pseudo label
  • Feature selection
• Concept drift handling
• Ensemble