关于「六格损失矩阵:三区决策的贝叶斯风险与两条切分点」,《三级灯号切分点》给出的定量结论是什么?
结论:估成真实值的 0.50 倍,即 25.0 元时,每百段多损失 17.5 元,相当于最优损失的 7.3 %;估成 2.00 倍,即 100.0 元时,多损失 23.0 元,即 9.6 %。估偏三成左右,代价只有两个百分点上下:估成 35.0 元多付 5.2 元,估成 70.0 元多付 5.4 元。只有偏到四倍,损失才变得显著:估成 12.5 元多付 73.4 元,即 30.7 %;估成 200.0 元多付 89.9 元,即 37.6 %。
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。
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: 0.50, 100.0, 12.5, 17.5, 2.00, 200.0
| LW_hat | factor | p_lo | p_hi | cost/100 | regret | regret % |
| 12.5 | 0.25 | 0.0000 | 0.8000 | 312.5 | 73.4 | 30.7 |
| 25.0 | 0.50 | 0.0000 | 0.9000 | 256.6 | 17.5 | 7.3 |
| 35.0 | 0.70 | 0.1429 | 0.9286 | 244.3 | 5.2 | 2.2 |
| 50.0 | 1.00 | 0.4000 | 0.9500 | 239.1 | 0.0 | 0.0 |
| 70.0 | 1.40 | 0.5714 | 0.9643 | 244.5 | 5.4 | 2.3 |
| 100.0 | 2.00 | 0.7000 | 0.9750 | 262.1 | 23.0 | 9.6 |
| 200.0 | 4.00 | 0.8500 | 0.9875 | 329.0 | 89.9 | 37.6 |
查看完整验算程序
# Chapter 2. p23: what if the owner misjudges the price of a wrong answer?
# Model (illustrative): the loss matrix of p21 (CA = 2, M = 0.8, CR = 8)
# with a true wrong-answer loss LW = 50. The owner sets both cuts from
# a guess LW_hat; the cost is then scored with the true LW. Regret is
# the extra yuan per 100 chats over the cuts set with the true value.
import math
LW, CA, M, CR = 50.0, 2.0, 0.8, 8.0
MU, SD = 3.0, 1.6
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 cuts(lw):
p_hi = 1 - CA / (M * lw)
p_lo = max(0.0, 1 - (CR - CA) / ((1 - M) * lw))
return p_lo, p_hi
def true_cost(lw_hat):
lo, hi = cuts(lw_hat)
terms = []
for p, w in PW:
if p >= hi:
r = (1 - p) * LW
elif p >= lo:
r = CA + (1 - p) * (1 - M) * LW
else:
r = CR
terms.append(w * r)
return 100 * math.fsum(terms)
base = true_cost(LW)
print("LW_hat factor p_lo p_hi cost/100 regret regret %")
for lw_hat in (12.5, 25.0, 35.0, 50.0, 70.0, 100.0, 200.0):
lo, hi = cuts(lw_hat)
c = true_cost(lw_hat)
print("%6.1f %6.2f %6.4f %6.4f %8.1f %6.1f %8.1f"
% (lw_hat, lw_hat / LW, lo, hi, c, c - base,
100 * (c - base) / base))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 2 章「六格损失矩阵:三区决策的贝叶斯风险与两条切分点」(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}
}