推定の偏りを測る ── 全体では当たっているのに、特定の群だけ外れる

import numpy as np
rng = np.random.default_rng(42)
# 群ごとの件数と、その群での的中率GROUPS = {    "A（多数派）": {"n": 9000, "acc": 0.95},    "B":           {"n":  800, "acc": 0.88},    "C（少数派）": {"n":  200, "acc": 0.52},}
total_correct = 0total_n = 0print(f"{'群':14s}{'件数':>7}{'的中率':>9}")for name, g in GROUPS.items():    correct = int(g["n"] * g["acc"])    total_correct += correct    total_n += g["n"]    print(f"{name:14s}{g['n']:7d}{g['acc']:9.1%}")
print(f"\n全体の的中率: {total_correct / total_n:.1%}")print("→ 全体は良好に見えるが、C群は当てずっぽうと大差ない")

import numpy as npimport pandas as pd

def accuracy_by_group(df, group_col, true_col, pred_col, min_n=30):    """群ごとの的中率を出す。件数が少ない群は別に示す。"""    rows = []    for name, g in df.groupby(group_col):        n = len(g)        acc = (g[true_col] == g[pred_col]).mean()        rows.append({            "群": name,            "件数": n,            "的中率": acc,            "判定": "件数不足" if n < min_n else "",        })
    out = pd.DataFrame(rows).sort_values("件数", ascending=False)
    # 全体との差を出す    overall = (df[true_col] == df[pred_col]).mean()    out["全体との差"] = out["的中率"] - overall    return out, overall

# 使い方# result, overall = accuracy_by_group(df, "region", "actual", "predicted")# print(f"全体: {overall:.1%}")# print(result.to_string(index=False))

import pandas as pd

def compare_distribution(df, true_col, pred_col):    """実際の分布と、推定結果の分布を比べる。"""    actual = df[true_col].value_counts(normalize=True).sort_index()    pred = df[pred_col].value_counts(normalize=True).sort_index()
    out = pd.DataFrame({"実際": actual, "推定": pred}).fillna(0)    out["差"] = out["推定"] - out["実際"]    return out.sort_values("差", key=abs, ascending=False)

# 推定が特定の値に寄っていれば、その値の「差」が大きく正になる# result = compare_distribution(df, "actual", "predicted")# print(result.to_string())