SuperModelingFactory 0.8.0__tar.gz → 0.8.2__tar.gz

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/ExcelMaster/ExcelMaster.py

@@ -241,7 +241,12 @@ class ExcelMaster(ExcelWorkbook):
         start_col = loc[1] if loc else self.curr_col
         written_range = [start_row, start_col, start_row + nrows - 1, start_col + ncols - 1]
         
-        worksheet.merge_range(*written_range, text, self.dict_cell_format[cformat])
+        if nrows == 1 and ncols == 1:
+            # xlsxwriter refuses to merge a single cell and writes nothing, so a
+            # one-column table used to lose its title; write the cell directly.
+            worksheet.write(start_row, start_col, text, self.dict_cell_format[cformat])
+        else:
+            worksheet.merge_range(*written_range, text, self.dict_cell_format[cformat])
 
         if self.verbose:
             logging.info(f"Merged Cells: {self.to_cell_range_text(*written_range)}")

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/ExcelMaster/Template.py

@@ -6,7 +6,6 @@ import numpy as np
 from datetime import datetime
 import pdb, re
 from PIL import Image
-pd.options.mode.chained_assignment = None
 
 from ExcelMaster.ExcelFormatTool import ExcelFormat
 from ExcelMaster.Utility import *

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/Binning_Tool.py

@@ -4,13 +4,8 @@ logger = logging.getLogger(__name__)
 import pandas as pd
 import numpy as np
 from scipy.stats import chi2_contingency, chi2
+from Modeling_Tool._utils.frames import as_binning_numeric
 
-# Available only in newer pandas versions. Older Airflow images should skip it.
-try:
-    pd.set_option('future.no_silent_downcasting', True)
-except (KeyError, ValueError):
-    pass
-pd.options.mode.chained_assignment = None  # default='warn'
 
 logging.basicConfig(level=logging.INFO, format="%(message)s")
 
@@ -540,7 +535,7 @@ def cre_pvt(df, var_name, tgt_name):
     >>> df_pvt = cre_pvt(df, var_name='income', tgt_name='default')
     """
     
-    df_pvt = df.groupby(var_name)[tgt_name].value_counts().unstack().fillna(0)
+    df_pvt = df.groupby(var_name, observed=False)[tgt_name].value_counts().unstack().fillna(0)
     df_pvt["n"] = df_pvt[0] + df_pvt[1]
     df_pvt["tr"] = df_pvt[1] / df_pvt["n"]
 
@@ -694,7 +689,9 @@ def get_bin_range(edges, precision = 5, ascending = False, left_sign = '(', righ
     
     i = 0
     reverse = not ascending
-    edges = sorted([round(x, precision) for x in edges], reverse = reverse)
+    # 极大边界（如 float 最大值）按精度取整会溢出成 inf，历来如此，不为此告警
+    with np.errstate(over="ignore"):
+        edges = sorted([round(x, precision) for x in edges], reverse = reverse)
     res = []
     while i < len(edges) - 1:
         left = edges[i]
@@ -708,6 +705,9 @@ def get_bin_range(edges, precision = 5, ascending = False, left_sign = '(', righ
 
 def _materialize_bin_columns(data, binned, bin_range_list, bin_num_col, bin_range_col):
     """Attach categorical bin numbers and labels without Python row callbacks."""
+    # The two bin columns are ours to add: work on our own frame so callers that
+    # pass a slice or their own DataFrame never get these columns written back.
+    data = data.copy(deep = False)
     codes = binned.cat.codes.to_numpy(dtype=np.intp, copy=False)
     range_values = np.empty(len(codes), dtype=object)
     range_values[:] = np.nan
@@ -935,7 +935,9 @@ def quick_binning(data, column, labels = None, nbins = 10, precision = 5, equal_
     >>> binned, edges = quick_binning(data, 'income', nbins=10, equal_freq=True)
     """
     
-    binning_series = data[column].round(precision)
+    # values near the float limits overflow when rounded, as they always did
+    with np.errstate(over="ignore"):
+        binning_series = as_binning_numeric(data[column]).round(precision)
     
     if include_missing:
         binning_series = binning_series.fillna(fillna)
@@ -986,14 +988,17 @@ def quick_binning(data, column, labels = None, nbins = 10, precision = 5, equal_
     if len(spec_values) > 0:
         fnl_breakpoints = sorted(list(set(list(fnl_breakpoints) + spec_values)))
         
-    binned, bin_edges = pd.cut(
-        binning_series, 
-        bins = fnl_breakpoints, 
-        labels = labels, 
-        right = right, 
-        include_lowest = include_lowest, 
-        retbins = True
-    )
+    # pandas rounds the edges to format interval labels; float-max edges overflow
+    # to inf there as they always did, so the numpy notice is not useful
+    with np.errstate(over="ignore"):
+        binned, bin_edges = pd.cut(
+            binning_series, 
+            bins = fnl_breakpoints, 
+            labels = labels, 
+            right = right, 
+            include_lowest = include_lowest, 
+            retbins = True
+        )
     
     orig_cat = [x for x in binned.cat.categories.tolist()]
     if ascending:

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/ODPS_Tool.py

@@ -6,14 +6,12 @@ import threading
 logger = logging.getLogger(__name__)
 import pandas as pd
 from odps import ODPS, options
-from odps.models import Schema, Column, Partition
-
-# Available only in newer pandas versions. Older Airflow images should skip it.
+from odps.models import Column, Partition
 try:
-    pd.set_option('future.no_silent_downcasting', True)
-except (KeyError, ValueError):
-    pass
-pd.options.mode.chained_assignment = None  # default='warn'
+    from odps.models import TableSchema as Schema
+except ImportError:  # older pyodps only ships the Schema name
+    from odps.models import Schema
+
 
 logging.basicConfig(level=logging.INFO, format="%(message)s")
 

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/Parallel_Engine.py

@@ -189,9 +189,15 @@ class ParallelApplyEngine:
         func_args: tuple[Any, ...],
         func_kwargs: dict[str, Any],
     ) -> None:
+        # joblib < 1.6 vendors cloudpickle; joblib >= 1.6 dropped the copy and depends on
+        # the cloudpickle package. Resolve the serializer outside the check below so an
+        # import problem is never reported as a non-serializable callable.
         try:
             from joblib.externals import cloudpickle
+        except ImportError:
+            import cloudpickle
 
+        try:
             cloudpickle.dumps((func, func_args, func_kwargs))
         except Exception as exc:
             raise TypeError(

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/Slope_Tool.py

@@ -1,12 +1,6 @@
 import logging
 import numpy as np
 import pandas as pd
-# Available only in newer pandas versions. Older Airflow images should skip it.
-try:
-    pd.set_option('future.no_silent_downcasting', True)
-except (KeyError, ValueError):
-    pass
-pd.options.mode.chained_assignment = None  # default='warn'
 logging.basicConfig(level=logging.INFO, format="%(message)s")
 
 

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/kDataFrame.py

@@ -5,12 +5,6 @@ from pandas import DataFrame
 from pandas import Series
 import numpy as np
 
-# Available only in newer pandas versions. Older Airflow images should skip it.
-try:
-    pd.set_option('future.no_silent_downcasting', True)
-except (KeyError, ValueError):
-    pass
-pd.options.mode.chained_assignment = None  # default='warn'
 
 logging.basicConfig(level=logging.INFO, format="%(message)s")
 

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Core/utils.py

@@ -1696,7 +1696,9 @@ def _calc_woe_iv_values(data, bad_pct, good_pct, fillwoe=True, filliv=True):
     if len(data[bad_pct]) > 0 and len(data[good_pct]) > 0:
         bad_values = data[bad_pct]
         good_values = data[good_pct]
-        woe = np.log(bad_values / good_values)
+        # 某一类占比为 0 的箱 WOE 为 ±inf（历来如此，调用方自行处理），不为此告警
+        with np.errstate(divide="ignore"):
+            woe = np.log(bad_values / good_values)
         iv = (bad_values - good_values) * woe
     else:
         woe = 0 if fillwoe else np.nan
@@ -1854,7 +1856,7 @@ def scoring(data, model, varlist, scr_name, keeplist = None, all_missing_spec_va
     nohit_condition = (pd.isnull(fnl_data[varlist]).sum(axis = 1) == len(varlist))
     if fnl_data[nohit_condition].shape[0] > 0:
         
-        all_missing_data = fnl_data[nohit_condition]
+        all_missing_data = fnl_data[nohit_condition].copy()
         other_data = fnl_data[~nohit_condition]
         
         

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Eval/Model_Eval_Tool.py

@@ -5,6 +5,7 @@ import numpy as np
 import pandas as pd
 from Modeling_Tool.Core.Binning_Tool import get_bin_range_list, super_binning
 from Modeling_Tool.Core.utils import load_model, calc_iv, calc_woe
+from Modeling_Tool._utils.frames import concat_non_empty
 from .evaluate_model import evaluate_performance
 from . import weighted_eval_utils as _weighted_eval
 
@@ -133,7 +134,9 @@ def _get_gains_table_scr(data, score, dep, nbins = 10, precision = 5,
     
 
     if add_func is not None:
-        gains_table_add = res.groupby(["_bin_num", "_bin_range"], dropna=False).apply(add_func)
+        # 显式选中全部列（含分组列）：add_func 照旧能看到 _bin_num / _bin_range，
+        # 也不触发 pandas 对 apply 默认包含分组列的弃用告警
+        gains_table_add = res.groupby(["_bin_num", "_bin_range"], dropna=False)[res.columns.unique().tolist()].apply(add_func)
         gains_table = gains_table.merge(gains_table_add, right_index = True, left_index = True, how = 'left')
     
     if retSummary:
@@ -257,6 +260,7 @@ def _get_gains_table_single(data, dep, nbins = 10, precision = 5, min_bin_prop =
         return -3
     
     if score is None:
+        data = data.copy()  # the scored column is ours; do not leave it on the caller's frame
         data['_mdl_scr'] = model.predict_proba(data.loc[:, varlist])[:, 1]
         score = '_mdl_scr'
         
@@ -510,7 +514,8 @@ def _get_perf_summary_single(train,
     oos_gains['index'] = 'oos'
     oot_gains['index'] = 'oot'
     
-    gains_summ = pd.concat([ins_gains, oos_gains, oot_gains])
+    # 缺失样本集的占位空表不参与拼接，否则整数列（N_BUMP / N_BINS）会被它变成 object
+    gains_summ = concat_non_empty([ins_gains, oos_gains, oot_gains])
     
     model_eval_result_df = model_eval_result_df.merge(gains_summ, on = ['index'], how = 'left')
 
@@ -705,6 +710,7 @@ def _get_cust_gains_table_single(data, dep, nbins = 10, precision = 5, min_bin_p
         return -3
     
     if score is None:
+        data = data.copy()  # the scored column is ours; do not leave it on the caller's frame
         data['_mdl_scr'] = model.predict_proba(data.loc[:, varlist])[:, 1]
         score = '_mdl_scr'
         
@@ -2609,7 +2615,7 @@ class PerformanceEvaluator:
                 gains_summ_list.append(gains_res)
 
             if gains_summ_list:
-                gains_summ = pd.concat(gains_summ_list)
+                gains_summ = concat_non_empty(gains_summ_list)
                 model_eval_result_df = model_eval_result_df.merge(gains_summ, on = ['index'], how = 'left')
 
             return model_eval_result_df

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Eval/evaluate_model.py

@@ -146,7 +146,11 @@ def summarize_pr(pr_df):
     pr_info: dict
         P-R曲线统计信息字典
     """
-    equalind = np.argmin(abs(pr_df['precision'] - pr_df['recall']))
+    gap = abs(pr_df['precision'] - pr_df['recall'])
+    if not gap.notna().any():
+        # 精确率或召回率全为 NaN（如样本只有一个类别）：平衡点无定义
+        return {'bep_index': np.nan, 'bep_threshold': np.nan, 'bep_precision': np.nan, 'bep_recall': np.nan}
+    equalind = np.argmin(gap)
     pr_info = {
         'bep_index': equalind, 
         'bep_threshold': pr_df['thresholds'][equalind], 
@@ -331,6 +335,10 @@ def summarize_roc(roc_df):
         return {'auc': np.nan, 'ks_index': np.nan, 'ks_threshold': np.nan, 'ks': np.nan}
     
     f = roc_df['tpr'] - roc_df['fpr']
+    if not f.notna().any():
+        # 只有一个类别时 TPR 或 FPR 全为 NaN：AUC、KS 及其阈值都无定义。以前 argmax
+        # 返回 -1 后按标签取 thresholds[-1] 抛 KeyError，整个评估失败
+        return {'auc': np.nan, 'ks_index': np.nan, 'ks_threshold': np.nan, 'ks': np.nan}
     roc_info = {
         'auc': auc(roc_df['fpr'], roc_df['tpr']),
         'ks_index': np.argmax(f),
@@ -388,11 +396,12 @@ def __plot_ks_axes(roc_df, ax):
     ax.plot(X, roc_df['tpr'], color=palette['ClassicBlueRedGrey'][1], label='True Positive Rate', linewidth=2)
 
     roc_info = summarize_roc(roc_df)
-    ks_vector = [
-        [X[roc_info['ks_index']], X[roc_info['ks_index']]], 
-        [roc_df['fpr'][roc_info['ks_index']], roc_df['tpr'][roc_info['ks_index']]],
-        ]
-    ax.plot(ks_vector[0], ks_vector[1], linewidth=4, color=palette['ClassicBlueRedGrey'][2], label='KS')
+    if not pd.isna(roc_info['ks_index']):
+        ks_vector = [
+            [X[roc_info['ks_index']], X[roc_info['ks_index']]],
+            [roc_df['fpr'][roc_info['ks_index']], roc_df['tpr'][roc_info['ks_index']]],
+            ]
+        ax.plot(ks_vector[0], ks_vector[1], linewidth=4, color=palette['ClassicBlueRedGrey'][2], label='KS')
     ax.set_title('Threshold={0:.3f}  KS={1:.3f}'.format(roc_info['ks_threshold'], roc_info['ks']), fontsize=15)
     ax.legend(loc=1, fontsize=12)
 
@@ -596,6 +605,36 @@ def _plot_weighted_kde_line(values, weights, ax, color, label):
         ax.plot(centers, density, color=color, linewidth=1.5, label=label)
 
 
+def _non_nan_1d(values):
+    """distplot 的输入处理：转 1 维 float 数组并去掉 NaN（inf 保留）。"""
+    values = np.asarray(values, dtype=float)
+    if values.ndim > 1:
+        values = values.squeeze()
+    return values[~np.isnan(values)]
+
+
+def _anchor_like_distplot(values, ax):
+    """distplot 未指定 color 时先画再删一个 (均值, 0) 点取默认颜色；这个点仍计入坐标轴
+    数据范围，使纵轴包含 0。照做一遍，保证只有 KDE 曲线的图（多模型、近似常数分数）
+    坐标范围不变。"""
+    anchor, = ax.plot(values.mean() if values.size else np.nan, 0)
+    anchor.remove()
+
+
+def _plot_score_hist(values, bins, ax, **hist_kws):
+    """等价于已弃用的 ``sns.distplot(values, bins=bins, hist=True, kde=False, hist_kws=...)``。"""
+    values = _non_nan_1d(values)
+    _anchor_like_distplot(values, ax)
+    ax.hist(values, bins, orientation="vertical", **hist_kws)
+
+
+def _plot_score_kde(values, bw_method, ax, color, label):
+    """等价于已弃用的 ``sns.distplot(values, hist=False, kde=True, kde_kws={'bw': ...})``。"""
+    values = _non_nan_1d(values)
+    _anchor_like_distplot(values, ax)
+    sns.kdeplot(x=values, ax=ax, color=color, label=label, bw_method=bw_method)
+
+
 def __plot_single_kde_axes(y_true, y_score, bins, ax, fontdicts, sample_weight=None):
     """在axes上绘制单个kde图.
     
@@ -615,12 +654,12 @@ def __plot_single_kde_axes(y_true, y_score, bins, ax, fontdicts, sample_weight=N
     __plot_kde_axes_base(ax, fontdicts)
 
     if sample_weight is None:
-        sns.distplot(y_score, bins=bins, hist=True, kde=False, ax=ax,
-                     hist_kws={'density': True, 'rwidth': 0.95, 'color': palette['ClassicBlueRedGrey'][2], 'alpha': 1, 'label': 'Total'}, )
-        sns.distplot(y_score[np.where(y_true==0)], bins=bins, hist=False, kde=True,
-                     kde_kws={'bw': 1/bins/2, 'color': palette['ClassicBlueRedGrey'][0], 'label': 'Neg KDE'}, )
-        sns.distplot(y_score[np.where(y_true==1)], hist=False, kde=True,
-                     kde_kws={'bw': 1/bins/2, 'color': palette['ClassicBlueRedGrey'][1], 'label': 'Pos KDE'}, )
+        _plot_score_hist(y_score, bins, ax, density=True, rwidth=0.95,
+                         color=palette['ClassicBlueRedGrey'][2], alpha=1, label='Total')
+        _plot_score_kde(y_score[np.where(y_true==0)], 1/bins/2, ax,
+                        color=palette['ClassicBlueRedGrey'][0], label='Neg KDE')
+        _plot_score_kde(y_score[np.where(y_true==1)], 1/bins/2, ax,
+                        color=palette['ClassicBlueRedGrey'][1], label='Pos KDE')
         true_mean = np.mean(y_true)
         score_mean = np.mean(y_score)
         title = 'N={0:,}  True={1:.2%}  Score={2:.2%}'.format(
@@ -687,8 +726,8 @@ def __plot_multi_kde_axes(y_true, y_score_dict, bins, ax, fontdicts):
     for i in range(len(models)):
         md = models[i]
         y_score = np.array(y_score_dict[md])
-        sns.distplot(y_score, bins=bins, hist=False, kde=True, ax=ax,
-                     kde_kws={'bw': 1/bins/2, 'color': palette['MorandiDark'][i], 'label': '{0} (Score={1:.2%})'.format(md, np.mean(y_score))}, )
+        _plot_score_kde(y_score, 1/bins/2, ax, color=palette['MorandiDark'][i],
+                        label='{0} (Score={1:.2%})'.format(md, np.mean(y_score)))
         ax.axvline(x=np.mean(y_score), linestyle='--', linewidth=1, color=palette['MorandiDark'][i])
     ax.axvline(x=np.mean(y_true), linestyle='-', linewidth=1, color=palette['ClassicGreyRed'][0], label='True')
 
@@ -720,7 +759,7 @@ def __agg(df):
         group_cols = ['y_group', 'thresholds']
     else:
         group_cols = ['thresholds']
-    df_agg = df.groupby(group_cols).agg(
+    df_agg = df.groupby(group_cols, observed=False).agg(
         min_score=pd.NamedAgg(column='y_score', aggfunc='min'),
         max_score=pd.NamedAgg(column='y_score', aggfunc='max'),
         n=pd.NamedAgg(column='y_true', aggfunc='count'),
@@ -987,7 +1026,9 @@ def calc_fixed_pct(y_true, y_score, y_group=None, bin_edges=None, ascending=True
     pct_df = __agg(df)
 
     avg_true = np.mean(y_true)
-    pct_df['lift'] = [x / avg_true for x in pct_df['cumavg_true']]
+    # 没有坏样本时 avg_true 为 0，lift 为 NaN / inf（结果照旧），不为此告警
+    with np.errstate(divide="ignore", invalid="ignore"):
+        pct_df['lift'] = [x / avg_true for x in pct_df['cumavg_true']]
     pct_df['gain'] = np.cumsum(pct_df['capture_rate'])
 
     return pct_df
@@ -1902,8 +1943,9 @@ def __evaluate_performance(y_true, y_score, nrow, ncol, i, dist_bins, pct_bins,
 
         pct_info = summarize_pct(pct_df, ascending=pct_ascending)
     
-    if len(y_true) < 2 or len(np.unique(y_true)) < 2:
-        # 返回默认性能指标（全部为NaN）
+    if len(y_true) < 2:
+        # 有效样本不足两行：返回默认性能指标（全部为NaN）。只有一个类别时照常汇总与出图，
+        # N / avgTrue / avgScore / 分位目标率都有定义，只有 KS / AUC 为 NaN
         return {
             'N': np.nan, 
             'avgTrue': np.nan, 

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Explainability/Model_Explainer.py

@@ -21,6 +21,7 @@ from __future__ import annotations
 
 import ast
 import importlib.metadata
+import inspect
 import sys
 import warnings
 
@@ -517,13 +518,20 @@ class ModelExplainer:
             table["importance_pct"] = table["mean_abs_shap"] / total if total else 0.0
         return table
 
-    def summary_plot(self, X=None, max_display=20, plot_type="dot", show=True, save_path=None):
-        """SHAP summary (beeswarm / bar) plot."""
+    def summary_plot(self, X=None, max_display=20, plot_type="dot", show=True, save_path=None, random_state=0):
+        """SHAP summary (beeswarm / bar) plot.
+
+        ``random_state`` seeds the point jitter through shap's ``rng`` argument
+        (shap >= 0.47), so the plot no longer depends on NumPy's global RNG;
+        older shap versions ignore it.
+        """
         shap = _lazy_shap()
         import matplotlib.pyplot as plt
 
         values, X_used = self._ensure_values(X)
         kwargs = {"max_display": max_display, "plot_type": plot_type, "show": False}
+        if "rng" in inspect.signature(shap.summary_plot).parameters:
+            kwargs["rng"] = np.random.default_rng(random_state)
         if not isinstance(X_used, pd.DataFrame) and self.feature_names is not None:
             kwargs["feature_names"] = self.feature_names
         shap.summary_plot(values, X_used, **kwargs)

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Feature/Feature_Screen.py

@@ -55,6 +55,7 @@ _MONOTONE_INIT_KEYS = frozenset({
     "sv_small_policy",
     "sv_woe_smoothing",
     "sv_smoothing_alpha",
+    "unseen_special_policy",
 })
 _MONOTONE_FIT_KEYS = frozenset({"chi2_binning", "chi2_p", "chi2_init_size", "n_jobs"})
 

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Model/Backward_Tool.py

@@ -24,7 +24,6 @@ import os
 import sys
 import copy
 import logging
-import warnings
 import functools
 from collections import OrderedDict
 from typing import Optional, List, Dict, Union, Any, Tuple
@@ -37,9 +36,6 @@ import matplotlib.ticker as mticker
 
 from Modeling_Tool.Core.sample_weight_utils import resolve_sample_weight
 
-# Suppress warnings for cleaner output
-warnings.filterwarnings('ignore')
-
 
 def _resolve_backward_perf_weight_col(
     split_name: str,

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Model/LRM_Tool.py

@@ -112,6 +112,19 @@ def lr_varimp(model):
     return varimp_df.sort_values('importance', ascending=False).reset_index(drop=True)
 
 
+def _predict_positive_proba(model, x_arr):
+    """P(y=1) for a design matrix given as a numpy array.
+
+    A model fitted on a DataFrame gets the same column names back (values are
+    taken by position, so predictions are unchanged); this avoids sklearn's
+    "X does not have valid feature names" warning.
+    """
+    names_in = getattr(model, 'feature_names_in_', None)
+    if names_in is not None and x_arr.ndim == 2 and x_arr.shape[1] == len(names_in):
+        x_arr = pd.DataFrame(x_arr, columns=names_in)
+    return model.predict_proba(x_arr)[:, 1]
+
+
 def fast_lr_pvalues(model, x, feature_names):
     """Coefficient p-values for a fitted sklearn LogisticRegression via the
     observed Fisher information — same formula as get_lr_statsmodel_summary
@@ -121,7 +134,7 @@ def fast_lr_pvalues(model, x, feature_names):
     from scipy import stats
 
     x_arr = x.values if hasattr(x, 'values') else np.array(x)
-    prob = model.predict_proba(x_arr)[:, 1]
+    prob = _predict_positive_proba(model, x_arr)
     w = prob * (1 - prob)
     X_design = np.hstack([np.ones((x_arr.shape[0], 1)), x_arr])
     fisher = X_design.T @ (X_design * w[:, None])
@@ -174,7 +187,7 @@ def get_lr_statsmodel_summary(model, x, y, feature_names=None):
     x_arr = x.values if hasattr(x, 'values') else np.array(x)
     y_arr = y.values if hasattr(y, 'values') else np.array(y)
 
-    prob = model.predict_proba(x_arr)[:, 1]
+    prob = _predict_positive_proba(model, x_arr)
     w = prob * (1 - prob)
     W = np.diag(w)
     X_design = np.hstack([np.ones((x_arr.shape[0], 1)), x_arr])
@@ -214,7 +227,7 @@ def _compute_log_likelihood(model, x, y, sample_weight=None):
     y_arr = y.values if hasattr(y, 'values') else np.array(y)
     weight = None if sample_weight is None else np.asarray(sample_weight, dtype=float)
 
-    prob = model.predict_proba(x_arr)[:, 1]
+    prob = _predict_positive_proba(model, x_arr)
     prob = np.clip(prob, 1e-15, 1 - 1e-15)
     point_ll = y_arr * np.log(prob) + (1 - y_arr) * np.log(1 - prob)
     if weight is None:
@@ -434,16 +447,19 @@ class FeatureSelectionAnalyzer:
 
         # Preserve the exact legacy call path for no weight and every
         # constant-weight vector.
-        if weight is None or bool(np.all(weight == weight[0])):
-            vif_values = [variance_inflation_factor(x, i) for i in range(x.shape[1])]
-        else:
-            from statsmodels.regression.linear_model import WLS
+        # A perfectly collinear feature has R² = 1 and VIF = inf; that is the
+        # expected result, not a floating-point problem worth a warning.
+        with np.errstate(divide="ignore"):
+            if weight is None or bool(np.all(weight == weight[0])):
+                vif_values = [variance_inflation_factor(x, i) for i in range(x.shape[1])]
+            else:
+                from statsmodels.regression.linear_model import WLS
 
-            vif_values = []
-            for i in range(x.shape[1]):
-                others = np.arange(x.shape[1]) != i
-                r_squared = WLS(x[:, i], x[:, others], weights=weight).fit().rsquared
-                vif_values.append(1.0 / (1.0 - r_squared))
+                vif_values = []
+                for i in range(x.shape[1]):
+                    others = np.arange(x.shape[1]) != i
+                    r_squared = WLS(x[:, i], x[:, others], weights=weight).fit().rsquared
+                    vif_values.append(1.0 / (1.0 - r_squared))
         vif_data = pd.DataFrame({
             'feature': work.columns,
             'VIF': vif_values,

{supermodelingfactory-0.8.0 → supermodelingfactory-0.8.2}/Modeling_Tool/Pipeline/credit_model.py

@@ -28,6 +28,19 @@ from ._common import (
 )
 
 
+def _any_column_has_value(frame: pd.DataFrame, columns: list[str], value: Any) -> bool:
+    """Whether any of ``columns`` in ``frame`` holds ``value`` (dtype-incompatible columns never do)."""
+    for column in columns:
+        if column not in frame.columns:
+            continue
+        try:
+            if bool(frame[column].eq(value).any()):
+                return True
+        except TypeError:
+            continue
+    return False
+
+
 @dataclass
 class CreditModelPipelineConfig:
     output_dir: str = "output"
@@ -76,6 +89,10 @@ class CreditModelPipelineConfig:
             "sv_smoothing_alpha": 0.0,
         }
     )
+    # unseen_special_policy (monotone only): declared special values absent from
+    # the fit sample — "normal_bin" (legacy) or "neutral" placeholder bins.
+    # Without an explicit "special_values" key the monotone self-fit declares the
+    # legacy -999999 sentinel only when the WOE fit sample contains it.
     monotone_woe_params: dict[str, Any] = field(
         default_factory=lambda: {
             "n_init_bins": 20,
@@ -85,6 +102,7 @@ class CreditModelPipelineConfig:
             "sv_small_policy": "keep",
             "sv_woe_smoothing": "none",
             "sv_smoothing_alpha": 0.0,
+            "unseen_special_policy": "normal_bin",
         }
     )
 
@@ -901,14 +919,15 @@ class CreditModelPipeline:
         )
 
         if cfg.woe_engine.lower() == "monotone":
-            params = merge_dict(
-                {
-                    "feature_cols": feature_cols,
-                    "target_col": cfg.target_col,
-                    "special_values": [-999999],
-                },
-                cfg.monotone_woe_params,
-            )
+            defaults = {"feature_cols": feature_cols, "target_col": cfg.target_col}
+            if "special_values" not in cfg.monotone_woe_params:
+                # 默认哨兵 -999999 只在拟合样本里真的出现时才声明：没出现时声明与否
+                # 分箱、打分完全一致，声明只会触发"声明了但没出现"告警（unseen_special_policy=
+                # 'neutral' 下还会给每个特征加占位箱）。显式传入的 special_values 不受影响。
+                defaults["special_values"] = (
+                    [-999999] if _any_column_has_value(fit_ins, feature_cols, -999999) else []
+                )
+            params = merge_dict(defaults, cfg.monotone_woe_params)
             # fit()-only kwargs must not reach MonotoneWOEBinner.__init__ —
             # n_jobs / chi2_p / chi2_init_size in monotone_woe_params used to
             # raise TypeError on this self-fit path (the screening-side