关于「在线贝叶斯校准:给每条规则一个会长大的贝塔分布」,《条件参数化规则蒸馏与在线校准》给出的定量结论是什么?

结论:第四章用贝塔后验与汤普森抽样在线校准,四万次会话的平均遗憾为 50.2;要检出付款率从 0.040 升到 0.044,每组需要 39,475 人。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书是非个人化经营分析,不构成投资、法律或医疗意见,不对应任何真实客户或成交。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.040, 0.044, 39475

A. minimum exposures for a 90% interval no wider than w
w0.020: n >= 992
w0.010: n >= 4,100
w0.005: n >= 16,571
B. sessions per arm to detect 0.040 -> p1 (alpha 0.05, 0.8)
p10.044 (lift 10%): 39,475
p10.048 (lift 20%): 10,317
p10.052 (lift 30%): 4,783
p10.060 (lift 50%): 1,863
C. true rate stays 0.040; chance a window looks <0.032
window100 sessions: 0.416
window300 sessions: 0.244
window 1,000 sessions0.113
window 3,000 sessions0.014
查看完整验算程序
# Chapter 4. p43: how many sessions before a rule's threshold may move?
# Part A: smallest exposure count n at which the 90% posterior interval
# of a rule with prior Beta(2, 48) and observed rate 0.04 is no wider
# than a preset ceiling w (k = round(0.04 n)).
# Part B: sessions per arm for a two-sided test (alpha 0.05, power 0.8)
# that a re-tuned rule moves the pay rate from 0.040 to p1.
# Part C: fixed-threshold edits made on small samples: how often does a
# rule whose true rate is unchanged look 20% worse by chance?
import math
import random
import statistics


def betacf(a, b, x):
    tiny = 1e-300
    qab, qap, qam = a + b, a + 1.0, a - 1.0
    c, d = 1.0, 1.0 - qab * x / qap
    d = 1.0 / (d if abs(d) > tiny else tiny)
    h = d
    for m in range(1, 400):
        m2 = 2 * m
        aa = m * (b - m) * x / ((qam + m2) * (a + m2))
        d = 1.0 + aa * d
        d = 1.0 / (d if abs(d) > tiny else tiny)
        c = 1.0 + aa / c
        c = c if abs(c) > tiny else tiny
        h *= d * c
        aa = -(a + m) * (qab + m) * x / ((a + m2) * (qap + m2))
        d = 1.0 + aa * d
        d = 1.0 / (d if abs(d) > tiny else tiny)
        c = 1.0 + aa / c
        c = c if abs(c) > tiny else tiny
        delta = d * c
        h *= delta
        if abs(delta - 1.0) < 1e-14:
            break
    return h


def ibeta(a, b, x):
    if x <= 0.0:
        return 0.0
    if x >= 1.0:
        return 1.0
    lb = (math.lgamma(a + b) - math.lgamma(a) - math.lgamma(b)
          + a * math.log(x) + b * math.log(1.0 - x))
    if x < (a + 1.0) / (a + b + 2.0):
        return math.exp(lb) * betacf(a, b, x) / a
    return 1.0 - math.exp(lb) * betacf(b, a, 1.0 - x) / b


def bquant(a, b, q):
    lo, hi = 0.0, 1.0
    for _ in range(80):
        mid = 0.5 * (lo + hi)
        if ibeta(a, b, mid) < q:
            lo = mid
        else:
            hi = mid
    return 0.5 * (lo + hi)


def width(n, p=0.04):
    k = round(p * n)
    a, b = 2.0 + k, 48.0 + n - k
    return bquant(a, b, 0.95) - bquant(a, b, 0.05)


def min_n(w):
    lo, hi = 1, 200000
    while lo < hi:
        mid = (lo + hi) // 2
        if width(mid) <= w:
            hi = mid
        else:
            lo = mid + 1
    return lo


def per_arm(p0, p1, alpha=0.05, power=0.8):
    z = statistics.NormalDist()
    za, zb = z.inv_cdf(1 - alpha / 2), z.inv_cdf(power)
    pb = (p0 + p1) / 2
    num = (za * math.sqrt(2 * pb * (1 - pb))
           + zb * math.sqrt(p0 * (1 - p0) + p1 * (1 - p1)))
    return math.ceil((num / (p1 - p0)) ** 2)


def main():
    print("A. minimum exposures for a 90% interval no wider than w")
    for w in (0.02, 0.01, 0.005):
        print("  w = {:.3f}: n >= {:,}".format(w, min_n(w)))
    print("B. sessions per arm to detect 0.040 -> p1 (alpha 0.05, 0.8)")
    for p1 in (0.044, 0.048, 0.052, 0.060):
        print("  p1 = {:.3f} (lift {:.0%}): {:,}".format(
            p1, p1 / 0.04 - 1, per_arm(0.040, p1)))
    print("C. true rate stays 0.040; chance a window looks <= 0.032")
    rng = random.Random(4301)
    for n in (100, 300, 1000, 3000):
        hit = 0
        for _ in range(4000):
            k = 0
            for _ in range(n):
                k += 1 if rng.random() < 0.04 else 0
            hit += 1 if k / n <= 0.032 else 0
        print("  window {:>5,} sessions: {:.3f}".format(n, hit / 4000))


if __name__ == "__main__":
    main()

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《高阶知识蒸馏进化、条件参数化编译与实盘转化反馈校准母典》第 4 章「在线贝叶斯校准:给每条规则一个会长大的贝塔分布」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-377 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026377,
  author    = {Fan, Yafu},
  title     = {Next-Generation Knowledge Distillation, Context-Parametric Compilation and Live-Conversion Feedback Calibration},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论