LightGBM 建模与监控¶

本教程使用确定性合成数据完成时间切分、LightGBM 调参、打分和当前期监控。每个变量都在 Notebook 中定义，不依赖其他页面或缓存输出。

前置条件与数据角色¶

运行前安装 mars-risk[ml,tuning]==0.0.24。development_df 用于建模，baseline_df 提供监控基准，monitoring_df 表示需要评估的当前数据。固定日期只用于复现时间切分，不代表特定业务月份。

In [ ]:

from datetime import date, timedelta

import numpy as np
import polars as pl

from mars.monitoring import MarsMonitor, generate_monitoring_alert
from mars.pipeline import MarsModelingPipeline, MarsModelingStep

SEED = 1206
ROW_COUNT = 240
rng = np.random.default_rng(SEED)
application_dates = [date(2025, 1, 1) + timedelta(days=index) for index in range(ROW_COUNT)]
income = rng.normal(5200, 1300, ROW_COUNT)
utilization = rng.uniform(0.05, 0.95, ROW_COUNT)
logit = -0.00045 * (income - 5200) + 3.2 * (utilization - 0.5)
probability = 1.0 / (1.0 + np.exp(-logit))

development_df = pl.DataFrame(
    {
        "apply_dt": application_dates,
        "period": [value.strftime("%Y-%m") for value in application_dates],
        "income": income,
        "utilization": utilization,
        "target": rng.binomial(1, probability),
    }
)
development_df.head()

from datetime import date, timedelta import numpy as np import polars as pl from mars.monitoring import MarsMonitor, generate_monitoring_alert from mars.pipeline import MarsModelingPipeline, MarsModelingStep SEED = 1206 ROW_COUNT = 240 rng = np.random.default_rng(SEED) application_dates = [date(2025, 1, 1) + timedelta(days=index) for index in range(ROW_COUNT)] income = rng.normal(5200, 1300, ROW_COUNT) utilization = rng.uniform(0.05, 0.95, ROW_COUNT) logit = -0.00045 * (income - 5200) + 3.2 * (utilization - 0.5) probability = 1.0 / (1.0 + np.exp(-logit)) development_df = pl.DataFrame( { "apply_dt": application_dates, "period": [value.strftime("%Y-%m") for value in application_dates], "income": income, "utilization": utilization, "target": rng.binomial(1, probability), } ) development_df.head()

训练并打分¶

Pipeline 按日期严格切分 train/val/oot，只执行一个轻量 trial，并关闭 artifact 落盘。

In [ ]:

pipeline = MarsModelingPipeline(
    target="target",
    features=["income", "utilization"],
    steps=[
        MarsModelingStep(
            name="modeling",
            model_type="lgb",
            time_col="apply_dt",
            split_ratios={"train": 0.6, "val": 0.2, "oot": 0.2},
            tune_params={
                "n_trials": 1,
                "startup_trials": 1,
                "num_boost_round": 20,
                "early_stopping_rounds": 5,
                "artifact_dir": None,
            },
        )
    ],
)

pipeline_result = pipeline.fit(development_df)
scored_df = pipeline.predict(development_df, pred_col="model_score")
pipeline_result.modeling_result.history_table.head()

pipeline = MarsModelingPipeline( target="target", features=["income", "utilization"], steps=[ MarsModelingStep( name="modeling", model_type="lgb", time_col="apply_dt", split_ratios={"train": 0.6, "val": 0.2, "oot": 0.2}, tune_params={ "n_trials": 1, "startup_trials": 1, "num_boost_round": 20, "early_stopping_rounds": 5, "artifact_dir": None, }, ) ], ) pipeline_result = pipeline.fit(development_df) scored_df = pipeline.predict(development_df, pred_col="model_score") pipeline_result.modeling_result.history_table.head()

构造基准期与当前期¶

较早样本作为 baseline_df；较新样本作为 monitoring_df，并将部分 target 置空以模拟尚未充分表现。

In [ ]:

baseline_df = scored_df.filter(pl.col("period") < "2025-06")
monitoring_df = (
    scored_df
    .filter(pl.col("period") >= "2025-06")
    .with_row_index("_row_index")
    .with_columns(
        pl.when(pl.col("_row_index") % 3 == 0)
        .then(None)
        .otherwise(pl.col("target"))
        .cast(pl.Int64)
        .alias("target")
    )
    .drop("_row_index")
)
monitoring_df.select("period", "target").head()

baseline_df = scored_df.filter(pl.col("period") < "2025-06") monitoring_df = ( scored_df .filter(pl.col("period") >= "2025-06") .with_row_index("_row_index") .with_columns( pl.when(pl.col("_row_index") % 3 == 0) .then(None) .otherwise(pl.col("target")) .cast(pl.Int64) .alias("target") ) .drop("_row_index") ) monitoring_df.select("period", "target").head()

In [ ]:

monitor_report = MarsMonitor(
    binner_params={"method": "quantile", "n_bins": 5},
    psi_include_missing=False,
).monitor(
    monitoring_df,
    features=["model_score", "income", "utilization"],
    target="target",
    benchmark_df=baseline_df,
    group_col="period",
    trend_column_order="asc",
)
monitor_report.summary_table.head()

monitor_report = MarsMonitor( binner_params={"method": "quantile", "n_bins": 5}, psi_include_missing=False, ).monitor( monitoring_df, features=["model_score", "income", "utilization"], target="target", benchmark_df=baseline_df, group_col="period", trend_column_order="asc", ) monitor_report.summary_table.head()

In [ ]:

monitor_report.target_observation_table

monitor_report.target_observation_table

In [ ]:

alert_text = generate_monitoring_alert(
    monitor_report,
    score_key="model_score",
    model_features=["income", "utilization"],
)
print(alert_text)

alert_text = generate_monitoring_alert( monitor_report, score_key="model_score", model_features=["income", "utilization"], ) print(alert_text)

结果边界¶

pipeline_result 保存建模步骤和调参结果；monitor_report 保存当前期分布、已表现样本指标和覆盖率。报警摘要只是结构化 report 的默认文本视图，实际阈值、通知和处置由调用方管理。