run-autogluon

Purpose

Provide a single, comprehensive reference for AutoGluon TabularPredictor usage in this repo, including model construction, training, evaluation, calibration, saving/loading, sklearn integration, and monotonic constraint workflows.

Usage

• “AutoGluon TabularPredictor arguments or workflow”
• “fit/predict_proba/fit_summary for AutoGluon”
• “calibrate or set decision threshold”
• “save/load AutoGluon models”
• “set best model in AutoGluon”
• “AutoGluon sklearn wrapper or PDP compatibility”
• “monotonic constraints in AutoGluon”

Instructions

1. Identify the task category:
   • Initialize predictor, fit, evaluate, infer, or persist.
   • Calibrate thresholds for binary classification.
   • Wrap predictor for sklearn usage.
   • Apply monotonic constraints.
2. Use the relevant section below to map arguments and steps.
3. Prefer minimal changes and align with course workflows (data in Data/, labs in Lectures/).

Core Workflow (Essentials)

1. Load data with TabularDataset.
2. Train with TabularPredictor(...).fit(...).
3. Evaluate with leaderboard() or fit_summary().
4. Predict with predict() or predict_proba().
5. Save/load using save() and load() as needed.

TabularPredictor Constructor (Key Args)

• label: target column name.
• problem_type: binary, multiclass, regression, quantile, or inferred.
• eval_metric: metric to optimize; defaults to AutoGluon choice by task.
• path: model artifact directory; defaults to AutogluonModels.
• verbosity: 0–4 for logging.
• log_to_file, log_file_path.
• sample_weight: column name or auto_weight/balance_weight.
• weight_evaluation: use weights during evaluation.
• groups: group-based splitting (experimental).
• positive_class: label for positive class in binary.
• Advanced: ignored_columns, cache_data, learner_type, trainer_type, default_base_path.

fit(...) (Key Args)

• train_data, tuning_data.
• time_limit, presets, hyperparameters.
• feature_metadata, feature_generator.
• Ensembling: auto_stack, num_bag_folds, num_bag_sets, num_stack_levels.
• fit_weighted_ensemble, full_weighted_ensemble_additionally.
• dynamic_stacking.
• Calibration: calibrate_decision_threshold, calibrate.
• Resources: num_cpus, num_gpus, memory_limit, fit_strategy.
• Deployment: refit_full, set_best_to_refit_full, keep_only_best, save_space.

fit_summary(...)

• verbosity: 0+; 2+ includes plots.
• show_plot: open plot in browser.
• Returns a summary dict for analysis/reporting.

predict_proba(...)

• Classification only.
• as_pandas: DataFrame vs ndarray.
• as_multiclass: for binary, return two columns vs positive class only.
• transform_features: set False only if data already transformed.

Threshold Calibration (Binary)

calibrate_decision_threshold(...)

• data: labeled calibration data or None for OOF/validation.
• metric: optimize target metric.
• decision_thresholds: int or list.
• secondary_decision_thresholds: optional fine search.
• subsample_size, verbose.

set_decision_threshold(...)

• decision_threshold: float in [0,1] for binary predictions.

set_model_best(...)

• model: model name from model_names()/leaderboard().
• save_trainer: persist default best across reloads.

save/load

• save(silent=True|False).
• load(path, verbosity=None, require_version_match=True, require_py_version_match=True, check_packages=True).
• Only load trusted artifacts (pickle).

sklearn Wrapper Integration

1. Implement ClassifierMixin + BaseEstimator (mixin should go first)
2. Track feature_names_, n_features_in_, classes_, is_fitted_.
3. In fit():
   • Convert X to DataFrame.
   • Add label column and optional sample_weight column.
   • Call TabularPredictor(...).fit(...).
4. In predict()/predict_proba():
   • Validate feature names/counts.
   • Convert to TabularDataset.
5. Use ./templates/wrapper_snippet.md as baseline layout.

Monotonic Constraints (Dictionaries)

1. Define a constraints dict: 1 non-decreasing, -1 non-increasing.
2. Build a feature-ordered list for GBM models.
3. Set hyperparameters:
   • GBM: list order.
   • CAT, XGB: dict form.
4. Retrain with updated hyperparameters.

Notes

• Prefer course data folders under Data/ and lab scripts under Lectures/.
• Keep changes aligned with existing lab workflows.