wildwood 0.2

scikit-learn compatible alternative random forests algorithms

WildWood

Scikit-Learn compatible Random Forest algorithms

Documentation | Reproduce experiments |

Installation

The easiest way to install wildwood is using pip

pip install wildwood

But you can also use the latest development from github directly with

pip install git+https://github.com/pyensemble/wildwood.git

Experiments

Experiments with hyperparameters optimization

To run experiments with hyperparameters optimization, under directory experiments/, use

python run_hyperopt_classfiers.py --clf_name WildWood --dataset_name adult

(with WildWood and on adult dataset in this example).

Some options are

• Setting --n_estimators or -t for number of estimators (for maximal number of boosting iterations in case of gradient boosting algorithms), default 100.
• Setting --hyperopt_evals or -n for number of hyperopt steps, default 50.

Experiments on default parameters

To run experiments with default parameters, under directory experiments/, use

python run_benchmark_default_params_classifiers.py --clf_name WildWood --dataset_name adult

(with WildWood and on adult dataset in this example).

Datasets and classifiers

For both run_hyperopt_classfiers.py and run_benchmark_default_params_classifiers. py, the available options for dataset_name are:

• adult
• bank
• breastcancer
• car
• cardio
• churn
• default-cb
• letter
• satimage
• sensorless
• spambase
• amazon
• covtype
• internet
• kick
• kddcup
• higgs

while the available options for clf_name are

• LGBMClassifier
• XGBClassifier
• CatBoostClassifier
• RandomForestClassifier
• HistGradientBoostingClassifier
• WildWood

Experiments presented in the paper

All the scripts allowing to reproduce the experiments from the paper are available in the experiments/ folder

1. Figure 1 is produced using fig_aggregation_effect.py.

2. Figure 2 is produced using n_tree_experiment.py.

3. Tables 1 and 3 from the paper are produced using run_hyperopt_classfiers.py with n_estimators=5000 for gradient boosting algorithms and with n_estimators=n for RFn and WWn

   • call

   python run_hyperopt_classfiers.py --clf_name <classifier> --dataset_name <dataset> --n_estimators <n_estimators>
   

   for each pair (<classifier>, <dataset>) to run hyperparameters optimization experiments;

   • use for example

   import pickle as pkl
   filename = 'exp_hyperopt_xxx.pickle'
   with open(filename, "rb") as f:
       results = pkl.load(f)
   df = results["results"]
   

   to retrieve experiments information, such as AUC, logloss and their standard deviation.

4. Tables 2 and 4 are produced using benchmark_default_params.py

   • call

   python run_benchmark_default_params_classifiers.py --clf_name <classifier> --dataset_name <dataset>
   

   for each pair (<classifier>, <dataset>) to run experiments with default parameters;

   • use similar commands to retrieve experiments information.

5. Using experiments results (AUC and fit time) done by run_hyperopt_classfiers.py, then concatenating dataframes and using fig_auc_fit_time.py to produce Figure 3.