parameters_info.md

Whitebox pipeline parameters:

General params:

- n_jobs
- debug

0) Simple typing and trash removal

0.0) Remove trash feats

    Medium:
    - th_nan
    - th_const

0.1) Typing (auto and user defined)

    Critical:
    - features_type (dict) {'age': 'real', 'education': 'cat', 'birth_date': (None, ("d", "wd"), ...}

0.2) Dates and categories encoding

    Critical:
    - features_type (for datetimes)

    Optional:
    - cat_alpha (int) - greater means more conservative encoding

1) Initial feature selection (selection based on gbm importance)

Critical:
- select_type (None or int)
- imp_type (if type(select_type) is int 'perm_imt'/'feature_imp')

Optional:
- imt_th (float) - threshold for select_type is None

2) Binning:

Critical:
- monotonic / features_monotone_constraints
- max_bin_count / max_bin_count
- min_bin_size

- cat_merge_to
- nan_merge_to

Medium:
- force_single_split

Optional:
- min_bin_mults
- min_gains_to_split

3) WoE estimation WoE = LN( ((% 0 in bin) / (% 0 in sample)) / ((% 1 in bin) / (% 1 in sample)) ):

Critical:
- oof_woe

Optional:
- woe_diff_th
- n_folds (if oof_woe)

4) Post selection:

4.0) Partial dependencies with target

Critical:
- auc_th

4.1) VIF

Critical:
- vif_th

4.2) Partial correlcations

Critical:
- pearson_th

5) Model based selection

Optional:
- n_folds
- l1_grid_size
- l1_exp_scale

6) Final model refit:

Critical:
- regularized_refit
- p_val (if not regularized_refit)
- validation (if not regularized_refit)

Optional:
- interpreted_model
- l1_grid_size (if regularized_refit)
- l1_exp_scale (if regularized_refit)

7) Report generation

- report_params