HyperGBM Versions Save

Add compatibility with scikit-learn v1.4.x Update experiment discriminator option to disable or enable it Make imbalanced-learn optional

• Add compatibility with lightgbm v4.x
• Fix bug: failed to pickle estimator on notebook, see: https://github.com/DataCanvasIO/HyperGBM/issues/104.

We add a few new features to this version:

• Multi-objectives optimization

  • optimization algorithm

    • add MOEA/D(Multiobjective Evolutionary Algorithm Based on Decomposition)
    • add Tchebycheff, Weighted Sum, Penalty-based boundary intersection approach(PBI) decompose approachs
    • add shuffle crossover, uniform crossover, single point crossover strategies for GA based algorithms
    • automatically normalize objectives of different dimensions
    • automatically convert maximization problem to minimization problem
    • add NSGA-II(Non-dominated Sorting Genetic Algorithm)
    • add R-NSGA-II(A new dominance relation for multicriteria decision making)

  • builtin objectives

    • number of features
    • prediction performance

• Add experiment option n_jobs
• Upgrade hboard to v0.1.1

• Fix bug: EarlyStopping does not work.

• Add compatibility with xgboost v1.6
• Add compatibility with cuML 22.08
• Add a shap explainer: HyperGBMShapExplainer
• Make python-geohash ( which requires native c compiler ) optional

• Add compatibility with scikit-learn v1.1
• Fix TfidfPrimitive

• Support custom metric in experiment visualization
• Set experiment cv default to False if eval_data is not None
• Fix issues: #80, #82

• Fix experiment report in gpu mode

• Add compatibility with cuML 22.02