关于「准确率的盲区:Chow 拒判规则与成本比」,《三级灯号切分点》给出的定量结论是什么?

结论:再往下看。答错代价升到人工的 8 倍,Chow 门限移到 0.8750,损失 389.2 元;A1 口径不动,损失 558.7 元;A2 口径的门限 0.8170 离最优不远,损失 402.8 元。到 16 倍,Chow 门限 0.9375,损失 514.8 元,A1 口径升到 1093.1 元,A2 口径 666.2 元,两种准确率口径中较差的那个多付 578.3 元。到 32 倍,最优门限 0.9688,A1 口径的损失 2161.8 元,是最优 623.4 元的三倍多。

数据口径:本书用纯标准库程序把命题写成可计算的模型,并修正蓝图三处:带拒识通道的最优识别早见于 Chow 1957 年的论文,1970 年的论文给出误差与拒判的一般关系;规则只对校准过的概率成立,把成本比算出的切分点直接套在过度自信机器人的自报分数上,每百段对话损失 265.4 元,等渗回归校准后 239.8 元;私域客服有「机器人起草、真人过目」的第三种动作,需要六格损失矩阵与两条切分点,示意参数下答对概率低于 0.4000 亮红灯、不低于 0.9500 亮绿灯。 书中参数为示意,对话全部虚构;全部计算由书中程序生成;程序不读写任何文件,不连接任何客服系统;本书是非个人化经营建议,任何线上灯号的改动须由主事者人工复核放行。

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: 0.8170, 0.8750, 0.9375, 0.9688, 1093.1, 16

bot accuracy if it answered everything: 0.897
A1 cut 0.500 (accuracy rule); A2 cut 0.8170 (answered acc 0.95)
LW/CRChow cutcost Chowcost A1cost A2worst regret
20.5000158.0158.0205.247.2
40.7500264.4291.6271.027.1
80.8750389.2558.7402.8169.5
160.9375514.81093.1666.2578.3
320.9688623.42161.81193.21538.5
查看完整验算程序
# Chapter 1. p11: a cut point tuned for accuracy ignores the cost ratio.
# Model (illustrative): a latent case difficulty z ~ N(3.0, 1.6); the
# bot answers correctly with calibrated probability p = 1 / (1 + e^-z).
# Two lights only: green (bot answers) or red (a human takes over at
# CR = 8 yuan). A wrong bot answer costs LW yuan. Chow's rule answers
# when p >= 1 - CR / LW. Two accuracy rules are shown for contrast:
# A1 answers when p >= 0.5 (best accuracy of a right/wrong predictor),
# A2 picks the lowest cut whose answered accuracy reaches 0.95.
# Exact quadrature on 4001 nodes; costs are yuan per 100 chats.
import math

MU, SD, CR = 3.0, 1.6, 8.0


def nodes(n=4001):
    zs = [MU + SD * (-8.0 + 16.0 * i / (n - 1)) for i in range(n)]
    ws = [math.exp(-0.5 * ((z - MU) / SD) ** 2) for z in zs]
    tot = math.fsum(ws)
    return [(1.0 / (1.0 + math.exp(-z)), w / tot) for z, w in zip(zs, ws)]


PW = nodes()


def cost(theta, lw):
    # expected loss per chat: answer -> (1 - p) LW, hand over -> CR
    terms = [w * ((1 - p) * lw if p >= theta else CR) for p, w in PW]
    return 100 * math.fsum(terms)


def answered_acc(theta):
    num = math.fsum(w * p for p, w in PW if p >= theta)
    den = math.fsum(w for p, w in PW if p >= theta)
    return num / den


acc = math.fsum(w * p for p, w in PW)
a2 = next(0.5 + k / 2000 for k in range(1000)
          if answered_acc(0.5 + k / 2000) >= 0.95)
print("bot accuracy if it answered everything: %.3f" % acc)
print("A1 cut 0.500 (accuracy rule); A2 cut %.4f (answered acc 0.95)"
      % a2)
print(" LW/CR  Chow cut  cost Chow  cost A1  cost A2  worst regret")
for ratio in (2, 4, 8, 16, 32):
    lw = ratio * CR
    t = 1 - CR / lw
    c0, c1, c2 = cost(t, lw), cost(0.5, lw), cost(a2, lw)
    print("%6d  %8.4f  %9.1f  %7.1f  %7.1f  %12.1f"
          % (ratio, t, c0, c1, c2, max(c1, c2) - c0))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《私域客服人机交接三级灯号精算、误判损失矩阵与最优置信度切分点母典》第 1 章「准确率的盲区:Chow 拒判规则与成本比」(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-266 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026266,
  author    = {Fan, Yafu},
  title     = {Three Lights for Handing Over between Bot and Human in Private Customer Service: Misclassification Loss Matrices, Chow's Reject Rule and Optimal Cut Points on Calibrated Confidence},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论