ReportML

1. 简介

ReportML 作为一款跨平台的机器学习框架，它不仅是支持快速建模的 AutoML 框架，还是一个自动输出模型报告的框架。考虑到很多业务场景中建模依旧还是使用 Excel 作为模型输出报告，ReportML 会记录自动建模中的各种细节，并产出一份非常美观的 Excel 模型报告。

1. SETUP2. 数据报告3. 变量WoE表4. 变量WoE图5. 变量逐月稳定性6. LR 模型结果7. LGBM 模型结果

2. 安装

ReportML 暂未开源，目前可提供加密后的安装包，用户下载后执行以下命令进行安装:

pip install ./report-1.0.0-py3-none-any.whl

3. 快速入门

目前 ReportML 暂未提供 CLI，后续会添加，使用需要自行构建 YAML 配置文件（目前配置项较多，后续会进行精简，更易于使用）：

使用 ReportML 快速输出模型报告

YAML 配置文件主程序

project_config:
    name: "uci_adult"
    version: 0.1
    created_date: ""
    updated_date: ""
    finished_date: "2024-06-20"
    author: "郁明敏"

    dataset_config:
    train_file: "data/preprocessed/uci_001_adult_dataset.csv"
    test_file: ""
    target: "target_income_greater_than_50k"
    date_columns:
        - dt
        - month
    include_columns: []
    test_from_train: true
    random_seed: 42
    test_size_from_train: 0.2
    val_size_from_train: 0.2
    test_split_by_date: "2024-04-01"
    feature_desc_dict:
        - feat_name: age
        feat_cate: "用户基础属性"
        feat_desc: "年龄"
        - feat_name: workclass
        feat_cate: "工作信息"
        feat_desc: "工作类别"
        - feat_name: education-num
        feat_cate: "用户基础属性"
        feat_desc: "教育等级"
        - feat_name: marital-status
        feat_cate: "用户基础属性"
        feat_desc: "婚姻状况"
        - feat_name: race
        feat_cate: "用户基础属性"
        feat_desc: "种族"
        - feat_name: sex
        feat_cate: "用户基础属性"
        feat_desc: "性别"
        - feat_name: capital-gain
        feat_cate: "其他"
        feat_desc: "未知"
        - feat_name: capital-loss
        feat_cate: "其他"
        feat_desc: "未知"
        - feat_name: hours-per-week
        feat_cate: "工作信息"
        feat_desc: "每周工作时长"
        - feat_name: native-country
        feat_cate: "用户基础属性"
        feat_desc: "祖国"
        - feat_name: occupation
        feat_cate: "工作信息"
        feat_desc: "职业"
    # 特征的数据字典文件，优先 feature_desc_dict 信息
    feature_desc_dict_file: "./feature_dict/uci001_adult_feat_dict.xlsx"

    # 报告类别: excel / web
    report_type: excel
    # 选用模型类别: lr / lgbm / xgb，可多选
    apply_models:
    - lr

    excel_config:
    output_path: results/uci_001_audit.xlsx


    eda_config:
    # 将 EDA 结果输出到 Excel 中的配置
    sheet_name: "2. 数据报告"
    cols_num_format:
        Missing_Rate: "0.00%"
    cols_databar:
        - Missing_Rate
        - Unique
    cols_font_color: {}
    exclude_cols: ["dt",  "month"]


    woe_config:
    # 将 WoE-IV 结果输出到 Excel 中的配置
    sheet_name: "3. 变量WoE"
    cols_num_format:
        Bad_Rate: "0.00%"
        Pct_Bad: "0.00%"
        Pct_Good: "0.00%"
        Pct_Bin: "0.00%"
        Bin_Bad_Total_Rate: "0.00%"
        Bin_Good_Total_Rate: "0.00%"
        Pct_Bad_Cum: "0.00%"
        Pct_Bin_Bad: "0.00%"
        Pct_Bin_Good: "0.00%"
    cols_databar:
        - Bad_Rate
        - WoE
    cols_font_color:
        Pct_Bad: "d76343"
        Pct_Good: "2596be"
    exclude_cols: ["dt",  "month"]
    apt_bins:
        - name: "age"
        bins: [0, 25, 30, 35]
    #  woe_conf_file: "conf/woe.json"
    #  woe_save_path: "conf/woe.json"


    select_config:
    # 如果 P1 和 P95 分位数相同，等同于 95% 的缺失率
    drop_missing_rate_threshold: 0.95
    drop_category_unique_num_threshold: 20
    drop_woe_bin_prop_threshold: 0.05
    drop_iv_threshold: 0.02
    drop_woe_not_monotonous: true
    # 相关系数大于 0.9，保留 IV 高的特征
    drop_corr_threshold: 0.9
    drop_psi_threshold: 0.1
    drop_lr_z_negative: true


    # 算法配置参数
    lr_config:
    sheet_name: "4. LR模型结果"
    exclude_cols: ["dt",  "month"]
    cols_num_format:
        Z_Prop: "0.00%"
        Shap: "0.00"
        Selected_Prop: "0.00%"
        Cum_Bad_Prop: "0.00%"
        Cum_Good_Prop: "0.00%"
        Bad_Rate: "0.00%"
        Bad_Prop: "0.00%"
        Good_Prop: "0.00%"
        Bin_Prop: "0.00%"
    cols_databar:
        - "z"
        - "P>|z|"
        - "IV"
        - "Corr"
        - "Shap"
        - "Selected_Cate_Rate"
        - "Selected_Model_Rate"
        - "Bad_Rate"
        - "KS"
        - "Z_Prop"
        - "Selected_Prop"
    cols_font_color:
        Cum_Bad_Prop: "d76343"

    epochs: 100
    learning_rate: 0.01


    lgbm_config:
    sheet_name: "5. LGBM模型结果"
    exclude_cols: ["dt",  "month"]
    cols_num_format:
        Shap: "0.00"
        Split: "0.00"
        Split_Prop: "0.00"
        Gain: "0.00"
        Gain_Prop: "0.00"
        Selected_Prop: "0.00%"
        Cum_Bad_Prop: "0.00%"
        Cum_Good_Prop: "0.00%"
        Bad_Rate: "0.00%"
        Bad_Prop: "0.00%"
        Good_Prop: "0.00%"
        Bin_Prop: "0.00%"

    cols_databar:
        - "IV"
        - "Gain"
        - "Gain_Prop"
        - "Split"
        - "Split_Prop"
        - "Corr"
        - "Shap"
        - "Shap_Prop"
        - "Selected_Cate_Rate"
        - "Selected_Model_Rate"
        - "Bad_Rate"
        - "KS"
        - "Selected_Prop"

    params:
        boosting: gbdt
        num_leaves: 10
        objective: binary
        max_depth: 4
        learning_rate: 0.05
        max_bin: 10
        max_cat_threshold: 5
        bagging_fraction: 0.8
        bagging_freq: 10
        feature_fraction: 0.8
        min_data_in_leaf: 500
        seed: 30
        metric:
        - auc
        - binary_logloss
        early_stopping_rounds: 100
        verbose: -1
    epochs: 3000
    save_path: ""
    model_path: ""

    xgb_config:
    epochs: 100
    learning_rate: 0.01

from reportml import ModelReport


model_report = ModelReport("uci001_adult.yml")
model_report.generate()