跳转至

v0.6.7

Patch release landing the remaining P1/P2 audit items from the defect review batch (G00, G02-G09, G17 selection/binning governance + G01 SAP split materialization). All new fields default to legacy behavior — default-config outputs are byte-identical to 0.6.6 (regression-asserted).

WOE Fit Ordering & Engine Reuse (G00)

  • woe_fit_scope — 'all' (legacy default) or 'post_missing_gate'. The gated survivor set feeds every WOE-consuming downstream stage (PSI/IVKS/corr/selection); gated features keep distribution diagnostics plus a woe_artifacts['missing_gate_dropped'] table.
  • Screens now attach their self-fitted monotone binner to WeightedScreenResult.woe_engine (+ woe_engine_meta); new woe_artifacts_from_screen_result wraps it into the CM _reuse_screening_woe contract and from_screen_result auto-calls it. WOE_Master engines are never attached (they hold train_data) — woe_table + meta + UserWarning only.

Selection Gates: IV Band, VIF, Group/Multi-Target, Truncation (G02-G06)

  • iv_upper_threshold (G02) on all four screen paths via shared _iv_band_keep; drops record reason='iv_above_upper' in the new dropped_detail frame (var/stage/metric/value/threshold/reason) alongside stage_tables.
  • G03/G04/G05 — new Feature/Screen_Gates.py: SelectionEvidence (lazy per-group / per-target IV+direction closures priced on post-corr survivors), gate order vif -> group_stability -> multi_target -> truncation, shared point_biserial_direction. FVP builds evidence from selection_group_dims; configured gates without evidence (CMP path) raise at feature_screen entry. Truncation is always last and never backfills.
  • G06 — compute_vif's statsmodels import now raises an actionable ImportError naming the optional extra (pip install "SuperModelingFactory[stats]"); the screen-side apply_vif_stage guards likewise before mutating anything.
  • All thresholds recorded in selection_config_snapshot.

LR P-Value Backward Elimination (G07)

  • lr_elimination_mode='pvalue' + lr_elimination_params (pvalue_threshold=0.05, min_features=1, max_iterations=20, tie_breaker). Post-fit loop drops the worst-p feature and refits until every coefficient clears the threshold; p-values come from the scipy Fisher-information summary of the exact shipped sklearn model.
  • New fast_lr_pvalues is the vectorized O(n·k) equivalent of get_lr_statsmodel_summary's p-value column (the legacy path builds an n×n diagonal weight matrix — unusable in loops at pipeline sample sizes).
  • Trace lands in feature_selection_summary['lr_elimination'] + lr_pvalue_elimination.csv; models['lr'] carries the reduced varlist so eval/explain follow. Mode/params validated in _validate_input; trace resets per _train_models call so pipeline instance reuse cannot leak a stale trace.

Binner Governance: Small Bins, Direction, Missing Strategy, Refine Floor (G08/G09/G17)

MonotoneWOEBinner gains eight opt-in governance params (all default None/'auto' = byte-identical legacy behavior):

  • G08 — min_bad_count / min_good_count / small_bin_policy (merge, warn, raise). merge collapses toward the WOE-closer neighbor recording merge_trace; warn/raise name feature, bin and counts.
  • G09 — monotone_direction (str or per-feature dict) / reference_target (point-biserial sign) / direction_conflict_policy (warn, raise, keep). Direction resolves before copy.copy(self) so parallel workers inherit it. New get_direction_summary() reports feat/direction/basis/is_monotonic.
  • Missing-bin semantics: missing_bin_strategy (empirical_special, fixed_woe, fail) with fit-time coherence checks.
  • G17 — refine_dtree(max_depth=) caps tree depth; refine_min_n_bins_policy (warn, enforce, raise) stops refinement from collapsing below min_n_bins (enforce keeps prefit bins).
  • Governance raises use BinningPolicyViolation(ValueError), piercing the historic per-feature fault tolerance; governance warns are UserWarning.

Verifiable Row-Level Split Materialization (G01)

  • SAP materialize_split=True + id_col replays each target's recommended (window, ratio, seed) through _materialize_one_split — the exact index generator extracted from candidate scoring, so statistics and materialized membership are identical by construction.
  • Output: row_level_split ([id_col, target_col, split_col_name]) + split_artifact (seed, ratio, oot_basis='window'|'cutoff', counts, sha256 ID hashes per segment, smf_version).
  • oot_cutoff overrides the trailing window (OOT = oot_time_dim >= cutoff). Loud integrity: duplicate ids raise with counts; segments are pairwise exclusive; coverage equals the mature-row set.
  • persist_split_map writes row_level_split.csv + split_artifact.json into output_paths. Materializing without id_col is rejected at validation.

GUI Field Metadata Sync

  • Labels for the selection gates (G02-G06), woe_fit_scope, selection_group_dims, lr_elimination_mode/params (G07), and the SAP materialization fields (G01).
  • New option enums: insufficient_group_policy (keep_warn, drop, raise), target_rules (all, any, min_pass_count).
  • monotone_woe_params nested fields for the eight binner governance params.

Behavior Notes

  • Byte-identical outputs under default configs (regression-asserted).
  • Announced default flips (next minor, 0.7.0): woe_fit_scope 'all'→'post_missing_gate' and refine_min_n_bins_policy None→'warn' — joining the already-announced gains_ascending None→True, synthesize_missing_oot None→False, search_eval_splits None→["oos"], and backward_report_splits None→[].

Validation

  • Full local pytest suite: 809 passed, 0 skipped (88 new governance/selection/split-materialization regressions on top of the 0.6.6-era 721 baseline).
  • CI green across the legacy/modern/bleeding × Python 3.11/3.12 matrix, source package checks, and version-consistency check.
  • Source package verification and wheel/sdist build passed (tasks.py verify).