easy-ml/feature/feature_utils.py

# -*- coding:utf-8 -*-
"""
@author: yq
@time: 2023/12/28
@desc:  特征工具类
"""
import numpy as np
import pandas as pd
import scorecardpy as sc
from statsmodels.stats.outliers_influence import variance_inflation_factor as vif

FORMAT_DICT = {
    # 比例类 -1 - 1
    "bin_rate1": np.arange(-1, 1 + 0.1, 0.1),

    # 次数类1 0 -10
    "bin_cnt1": np.arange(0, 11, 1),
    # 次数类2 0 - 20
    "bin_cnt2": [0, 1, 2, 3, 4, 5, 8, 10, 15, 20],
    # 次数类3 0 - 50
    "bin_cnt3": [0, 2, 4, 6, 8, 10, 15, 20, 25, 30, 35, 40, 45, 50],
    # 次数类4 0 - 100
    "bin_cnt4": [0, 3, 6, 10, 15, 20, 30, 40, 50, 100],

    # 金额类1 0 - 1w
    "bin_amt1": np.arange(0, 1.1e4, 1e3),
    # 金额类2 0 - 5w
    "bin_amt2": np.arange(0, 5.5e4, 5e3),
    # 金额类3 0 - 10w
    "bin_amt3": np.arange(0, 11e4, 1e4),
    # 金额类4 0 - 20w
    "bin_amt4": [0, 1e4, 2e4, 3e4, 4e4, 5e4, 8e4, 10e4, 15e4, 20e4],
    # 金额类5 0 - 100w
    "bin_amt5": [0, 5e4, 10e4, 15e4, 20e4, 25e4, 30e4, 40e4, 50e4, 100e4],

    # 年龄类
    "bin_age": [20, 25, 30, 35, 40, 45, 50, 55, 60, 65],
}


# 粗分箱
def f_format_bin(data_describe: pd.Series, raw_v):
    percent10 = data_describe["10%"]
    percent90 = data_describe["90%"]
    format_v = raw_v

    # 筛选最合适的标准化分箱节点
    bin = None
    for k, v_list in FORMAT_DICT.items():
        bin_min = min(v_list)
        bin_max = max(v_list)
        if percent10 >= bin_min and percent90 <= bin_max:
            if bin is None:
                bin = (k, bin_max)
            elif bin[1] > bin_max:
                bin = (k, bin_max)

    if bin is None:
        return format_v

    # 选择分箱内适合的切分点
    v_list = FORMAT_DICT[bin[0]]
    for idx in range(1, len(v_list)):
        v_left = v_list[idx - 1]
        v_right = v_list[idx]
        # 就近原则
        if v_left <= raw_v <= v_right:
            format_v = v_right if (raw_v - v_left) - (v_right - raw_v) > 0 else v_left
    if format_v not in v_list:
        if format_v > v_list[-1]:
            format_v = v_list[-1]
        if format_v < v_list[0]:
            format_v = v_list[0]

    return format_v


# 此函数判断list的单调性，允许至多N次符号变化
def f_judge_monto(bd_list: list, pos_neg_cnt: int = 1) -> int:
    start_tr = bd_list[1] - bd_list[0]
    tmp_len = len(bd_list)
    pos_neg_flag = 0
    for i in range(2, tmp_len):
        tmp_tr = bd_list[i] - bd_list[i - 1]
        # 后一位bad_rate减前一位bad_rate，保证bad_rate的单调性
        # 记录符号变化, 允许 最多一次符号变化，即U型分布
        if (tmp_tr >= 0 and start_tr >= 0) or (tmp_tr <= 0 and start_tr <= 0):
            # 满足趋势保持，查看下一位
            continue
        else:
            # 记录一次符号变化
            start_tr = tmp_tr
            pos_neg_flag += 1
            if pos_neg_flag > pos_neg_cnt:
                return False
    # 记录满足趋势要求的变量
    if pos_neg_flag <= pos_neg_cnt:
        return True
    return False


def f_get_corr(data: pd.DataFrame, meth: str = 'spearman') -> pd.DataFrame:
    return data.corr(method=meth)


def f_get_ivf(data: pd.DataFrame) -> pd.DataFrame:
    if len(data.columns.to_list()) <= 1:
        return None
    vif_v = [vif(data.values, data.columns.get_loc(i)) for i in data.columns]
    vif_df = pd.DataFrame()
    vif_df["变量"] = data.columns
    vif_df['vif'] = vif_v
    return vif_df


def f_calcu_model_ks(data, y_column, sort_ascending):
    var_ks = data.groupby('MODEL_SCORE_BIN')[y_column].agg([len, np.sum]).sort_index(ascending=sort_ascending)
    var_ks.columns = ['样本数', '坏样本数']
    var_ks['好样本数'] = var_ks['样本数'] - var_ks['坏样本数']
    var_ks['坏样本比例'] = (var_ks['坏样本数'] / var_ks['样本数']).round(4)
    var_ks['样本数比例'] = (var_ks['样本数'] / var_ks['样本数'].sum()).round(4)
    var_ks['总坏样本数'] = var_ks['坏样本数'].sum()
    var_ks['总好样本数'] = var_ks['好样本数'].sum()
    var_ks['平均坏样本率'] = (var_ks['总坏样本数'] / var_ks['样本数'].sum()).round(4)
    var_ks['累计坏样本数'] = var_ks['坏样本数'].cumsum()
    var_ks['累计好样本数'] = var_ks['好样本数'].cumsum()
    var_ks['累计样本数'] = var_ks['样本数'].cumsum()
    var_ks['累计坏样本比例'] = (var_ks['累计坏样本数'] / var_ks['总坏样本数']).round(4)
    var_ks['累计好样本比例'] = (var_ks['累计好样本数'] / var_ks['总好样本数']).round(4)
    var_ks['KS'] = (var_ks['累计坏样本比例'] - var_ks['累计好样本比例']).round(4)
    var_ks['LIFT'] = ((var_ks['累计坏样本数'] / var_ks['累计样本数']) / var_ks['平均坏样本率']).round(4)
    return var_ks.reset_index()


def f_get_model_score_bin(df, card, bins=None):
    train_score = sc.scorecard_ply(df, card, print_step=0)
    df['score'] = train_score
    if bins is None:
        _, bins = pd.qcut(df['score'], q=10, retbins=True, duplicates="drop")
        bins = list(bins)
        bins[0] = -np.inf
        bins[-1] = np.inf
    score_bins = pd.cut(df['score'], bins=bins)
    df['MODEL_SCORE_BIN'] = score_bins.astype(str).values
    return df, bins


def f_calcu_model_psi(df_train, df_test):
    tmp1 = df_train.groupby('MODEL_SCORE_BIN')['MODEL_SCORE_BIN'].agg(['count']).sort_index(ascending=True)
    tmp1['样本数比例'] = (tmp1['count'] / tmp1['count'].sum()).round(4)
    tmp2 = df_test.groupby('MODEL_SCORE_BIN')['MODEL_SCORE_BIN'].agg(['count']).sort_index(ascending=True)
    tmp2['样本数比例'] = (tmp2['count'] / tmp2['count'].sum()).round(4)
    psi = ((tmp1['样本数比例'] - tmp2['样本数比例']) * np.log(tmp1['样本数比例'] / tmp2['样本数比例'])).round(4)
    psi = psi.reset_index()
    psi = psi.rename(columns={"样本数比例": "psi"})
    psi['训练样本数'] = list(tmp1['count'])
    psi['测试样本数'] = list(tmp2['count'])
    psi['训练样本数比例'] = list(tmp1['样本数比例'])
    psi['测试样本数比例'] = list(tmp2['样本数比例'])
    return psi