Home Credit 外部来源分交叉与保存后查询¶

使用本批全量公开统计。EXT_SOURCE_2/3 是官方外部来源分，TARGET 是原始标签，APPLY_DATE 为模拟日期。这里只加载快照，不读取原 CSV。完整复算见 Home Credit 输入指南。

In [ ]:

from pathlib import Path
from mars.reporting import load_report

root = Path.cwd()
while not (root / "docs/assets/cases").is_dir() and root.parent != root:
    root = root.parent
assets = root / "docs/assets/cases"
report = load_report(assets / "score-cross.marsreport")
report.describe()

from pathlib import Path from mars.reporting import load_report root = Path.cwd() while not (root / "docs/assets/cases").is_dir() and root.parent != root: root = root.parent assets = root / "docs/assets/cases" report = load_report(assets / "score-cross.marsreport") report.describe()

In [ ]:

page = report.query_page("cells", filters={"target": "TARGET", "group": "1val"}, limit=3)
print(page["reference"])
page["data"]

page = report.query_page("cells", filters={"target": "TARGET", "group": "1val"}, limit=3) print(page["reference"]) page["data"]

In [ ]:

unlabeled = report.query_page("overall", filters={"target": "TARGET", "group": "4_oot_undue"}, limit=2)
unlabeled["data"]

unlabeled = report.query_page("overall", filters={"target": "TARGET", "group": "4_oot_undue"}, limit=2) unlabeled["data"]

无标签组保留人数与分布，Bad Rate 等监督指标按实际未观测状态解释。切点来自 0train，其他分区复用。分组日期不能证明真实 OOT。

In [ ]:

from mars.reporting import get_correlation_matrix

correlation = load_report(assets / "raw-correlation.marsreport")
print(correlation.query_page("pairs", limit=3)["data"])
get_correlation_matrix(correlation, features=["AMT_INCOME_TOTAL", "AMT_ANNUITY"])

from mars.reporting import get_correlation_matrix correlation = load_report(assets / "raw-correlation.marsreport") print(correlation.query_page("pairs", limit=3)["data"]) get_correlation_matrix(correlation, features=["AMT_INCOME_TOTAL", "AMT_ANNUITY"])

In [ ]:

binning = load_report(assets / "binning.marsreport")
print(binning.query_page("summary", features="EXT_SOURCE_2", limit=1)["data"])
print(binning.to_ai_context(tables=["summary"], max_chars=16000))

binning = load_report(assets / "binning.marsreport") print(binning.query_page("summary", features="EXT_SOURCE_2", limit=1)["data"]) print(binning.to_ai_context(tables=["summary"], max_chars=16000))