Upgrading from 0.8.2 to 0.9.3

0.9.1 is the release that follows 0.8.2 on PyPI (0.9.0 was prepared but never published), then come 0.9.2 and 0.9.3. They change some defaults and fix defects that change numbers. This page lists, in one place, what to pass to get the 0.8.2, 0.9.0, 0.9.1 or 0.9.2 behavior back where a switch exists, and which changes are bug fixes without one. The details are in the v0.9.0, v0.9.1, v0.9.2 and v0.9.3 notes. Upgrading from 0.9.2 to 0.9.3 covers projects that rerun a screening.

Changed Defaults

Setting	0.8.2	0.9.0	0.9.1	0.9.2	0.9.3	Keep the old behavior with
MonotoneWOEBinner(unseen_special_policy=...)	"normal_bin"	"neutral"	"neutral"	"neutral"	"neutral"	unseen_special_policy="normal_bin"
MonotoneWOEBinner(sv_total_basis=...)	(did not exist; behaved as "ordinary")	"all"	"all"	"all"	"all"	sv_total_basis="ordinary"
CreditModelPipelineConfig(warm_start_score_scope=...)	(did not exist; behaved as "train")	"full"	"full"	"full"	"full"	warm_start_score_scope="train"
MonotoneWOEBinner(min_bad_count=..., min_good_count=..., small_bin_policy=...)	None, None, None	None, None, None	1, 1, "merge"	1, 1, "merge"	1, 1, "merge"	small_bin_policy=None
Special-value bins below min_bad_count / min_good_count (class-pure with the defaults)	Not checked	Not checked	Reported: UserWarning, sv_class_pure column, _sv_pure_stats; numbers unchanged with the default sv_small_policy="keep", neutralized or merged with the opt-in "neutral" / "merge_missing"	As in 0.9.1	As in 0.9.1	min_bad_count=None, min_good_count=None (also stops the merging of class-pure ordinary bins), or filter the warning
ODPS_PROJECT, ODPS_ENDPOINT for ODPSRunner	Optional, with built-in defaults	Optional, with built-in defaults	Required (KeyError if unset)	Required	Required	Set both variables
screen_iv_convention of FeatureScreenConfig, FeatureValidationPipelineConfig, CreditModelPipelineConfig, VarExtractionInsights and CorrelationFilter: the screening IV on WOE bins	(did not exist; behaved as "raw")	(as "raw")	(as "raw")	"guarded": class-pure bins are left out of the IV sum, which changes the IV, iv_threshold, the correlation decisions and the truncation ranking of features with such bins	As in 0.9.2	screen_iv_convention="raw" (bit-for-bit the 0.9.1 IV tables and selections)
iv_upper_threshold (leakage limit) of FeatureScreenConfig, screen_config_from_mapping, weighted_feature_screen, the FVP selection_params and the CMP feature_selection	None (off)	None	None	None	2.0; on WOE bins under "guarded" it tests the larger of iv_raw and the floored IV, never the guarded iv	iv_upper_threshold=None (or your previous explicit value)
Leak warning: UserWarning and iv_gap_flag when iv_raw / iv >= 2 or iv_raw - iv >= 0.5	(did not exist)	(did not exist)	(did not exist)	(did not exist)	On (iv_gap_ratio_warn=2.0, iv_gap_abs_warn=0.5); warns and flags only	Filter the warning, or set both thresholds to inf
on_empty_stage="keep_all_warn" (the default): a stage that would drop every feature	Keeps every feature	As in 0.8.2	As in 0.8.2	As in 0.8.2	Never brings back a feature that iv_upper_threshold dropped; from the IV stage on keeps the fallback_top_k (5) features with the highest screening IV (fallback_kept in dropped_detail); raises EmptyStageError when only leak drops are left; the missing-rate and PSI stages still keep every feature	on_empty_stage="keep_all" (new value; for comparisons with old results only)
Leak-gate warning: a UserWarning per feature that iv_upper_threshold drops	(did not exist)	(did not exist)	(did not exist)	(did not exist)	On whenever the limit drops a feature ("possible target leakage; check the field's time point (PIT) and source before training")	Filter the warning, or switch the limit off

The three MonotoneWOEBinner settings are also keys of monotone_woe_params in CreditModelPipelineConfig and FeatureValidationPipelineConfig (and of feature_screen). A dict passed as monotone_woe_params replaces the default dict, so start from the default when you add the legacy keys:

from Modeling_Tool import CreditModelPipelineConfig

legacy_woe = dict(
    CreditModelPipelineConfig().monotone_woe_params,
    unseen_special_policy="normal_bin",
    sv_total_basis="ordinary",
    small_bin_policy=None,
)
config = CreditModelPipelineConfig(monotone_woe_params=legacy_woe, warm_start_score_scope="train")
print(config.monotone_woe_params["small_bin_policy"], config.warm_start_score_scope)

The screening IV of 0.9.1 comes back with one setting per object; it does not touch the binning:

from Modeling_Tool import CreditModelPipelineConfig, FeatureScreenConfig, FeatureValidationPipelineConfig

screen = FeatureScreenConfig(iv_use_woe_bins=True, screen_iv_convention="raw")
fvp = FeatureValidationPipelineConfig(screen_iv_convention="raw")
cmp = CreditModelPipelineConfig(screen_iv_convention="raw")
print(screen.screen_iv_convention, fvp.screen_iv_convention, cmp.screen_iv_convention)

from Modeling_Tool import MonotoneWOEBinner

binner = MonotoneWOEBinner(
    feature_cols=["score"],
    target_col="is_bad",
    unseen_special_policy="normal_bin",
    sv_total_basis="ordinary",
    small_bin_policy=None,
)
print(binner.small_bin_policy, binner.sv_total_basis)

Saved Objects

• Scorecards, WOE tables and models built with an earlier version keep their numbers until they are refitted.
• A pickled MonotoneWOEBinner keeps the settings it was created with when it is loaded, applied or refitted. A binner from 0.8.2 or earlier keeps unseen_special_policy="normal_bin", sv_total_basis="ordinary" and small_bin_policy=None; a binner from 0.9.0 keeps small_bin_policy=None.
• Bins loaded with load_woe_bins keep the WOE they were saved with.
• With the default min_bin_size=0.03, small_bin_policy="merge" also merges an ordinary bin under 3% of the rows; in the default flow such a bin only comes from tied values at the equal-frequency edges. min_bin_size=0 restricts the merging to class-pure bins.
• The .pyi type stubs of 0.9.1 show the current signatures and defaults (they were stale up to 0.9.0).
• screen_iv_convention (0.9.2) changes the screening only: binners, WOE tables and apply_woe give the same values as in 0.9.1, so existing models score identically. When you build a new model, resolve every class-pure bin (flagged sv_class_pure, or listed in the screening's iv_sv_pure_bins) before training: its WOE is set by eps=1e-6 and has no statistical meaning.

Upgrading from 0.9.2 to 0.9.3

0.9.3 turns the leakage limit on (iv_upper_threshold=2.0) and adds a leak warning, so a rerun of the screening can drop features that 0.9.2 kept: those whose floored IV or iv_raw is above 2.0. A dropped leak is never brought back by an empty-stage fallback, and that fallback now keeps at most fallback_top_k (5) features from the IV stage on, so a stage that every feature fails can keep fewer features than in 0.9.2, or stop with EmptyStageError when only leak drops are left.

Projects that are mid-flight or already frozen

If you rerun the screening of a project whose variables are already chosen, pass screen_iv_convention="raw" and iv_upper_threshold=None (or the explicit value you used before) to keep the same selected variables. Under "raw" the limit tests the same value as in 0.9.2, so a previous explicit value gives the previous drops. If a stage of that screening ever dropped every feature (a <stage>_fallback row in its summary), also pass on_empty_stage="keep_all", the fallback of 0.9.2; it changes nothing otherwise, and it is meant only for reproducing old results. The warnings may still appear; they change nothing.

from Modeling_Tool import CreditModelPipelineConfig, FeatureScreenConfig, FeatureValidationPipelineConfig

frozen_screen = FeatureScreenConfig(
    iv_use_woe_bins=True, screen_iv_convention="raw", iv_upper_threshold=None, on_empty_stage="keep_all",
)
frozen_fvp = FeatureValidationPipelineConfig(
    screen_iv_convention="raw", selection_params={"iv_upper_threshold": None, "on_empty_stage": "keep_all"},
)
frozen_cmp = CreditModelPipelineConfig(
    screen_iv_convention="raw",
    feature_selection=dict(
        CreditModelPipelineConfig().feature_selection, iv_upper_threshold=None, on_empty_stage="keep_all",
    ),
)
print(frozen_screen.on_empty_stage, frozen_fvp.selection_params, frozen_cmp.feature_selection["iv_upper_threshold"])

New projects should keep the defaults and review every feature that the warnings name or that dropped_detail lists with iv_above_upper / iv_above_upper_sv_pure (check its time point and source), and every feature marked fallback_kept. An EmptyStageError that names only leak drops means that nothing else is left to screen: fix the data or the feature list rather than switching the fallback back.

New Outputs (0.9.2)

These add information without changing a number of 0.9.1:

Output	What it adds
iv_table, the iv table of the selection summaries, ivks_summary, the WOE-engine report of VarExtractionInsights	iv (under screen_iv_convention), iv_raw (every bin counted, the 0.9.1 value), iv_sv_pure and n_sv_pure (the IV and rows of the class-pure bins). iv_table keeps iv_weighted, equal to iv
stage_tables["iv_sv_pure_bins"], FeatureValidationPipelineResult.ivks_sv_pure_bins, feature_selection_summary["iv_sv_pure_bins"], VarExtractionInsights.iv_sv_pure_bins	The class-pure special-value bins, with n < 30 marked low_count, as candidate rules for review
dropped_detail of the leakage gate (iv_upper_threshold)	The reason "iv_above_upper_sv_pure" when a feature exceeds the limit only because of its class-pure special-value bins; the gate drops the same features as in 0.9.1
iv_gap_flag in iv_table and the iv tables of the selection summaries (0.9.3)	True when iv_raw is far above iv (the leak warning); CreditModelPipeline adds dropped_detail to feature_selection_summary
dropped_detail reason fallback_kept and FeatureScreenConfig.fallback_top_k (0.9.3)	The features that an empty-stage fallback kept; feature_selection_summary["hard_dropped"] lists the leak drops that a failed screening of CreditModelPipeline keeps out, and the FVP config_snapshot records fallback_top_k

Fixes Without a Legacy Switch

These changes correct results that were wrong; no setting brings the old numbers back.

Area	What changes
Equal-frequency binning with include_missing=True (WOE_Master, get_woe_table, WOETransformer, the non-monotone woe_engine, unweighted Gains and cross-risk tables)	Missing values get their own bin instead of sharing the lowest bin with real values; the default fill value -1.797e308 no longer overflows to -inf
RejectInferencePipeline	Rows without a label are no longer trained as goods; ri_score_direction="high_good" with a pre-score trained by the pipeline raises instead of inverting the reject labels; the random OOT is held out of the pre-score, so OOT labels no longer reach the RI models through the inferred reject labels; an external reference is rescored with the pipeline's pre-score
ScoreComparisonPipeline, model_perf_compare	Scores are compared on the same rows (perf_common_rows=False evaluates each score on its own rows, which reproduces 0.9.0 with one comparison score); numeric group columns are evaluated; the weighted Gains table no longer counts unlabelled rows as goods or ranks missing scores; a group with exactly min_data_size rows is kept
Everything listed in the v0.9.0 notes, sections 2, 4 and 5	See the "Results that change" boxes there