关于「红黄绿灯:ROC 曲线上的代价敏感移交门限」,《私域等待、转人工门限与客诉阻断》给出的定量结论是什么?

结论:第四章说明红黄绿灯门限必须随毛利移动,固定在中点的门限对高价商品的代价是 1232.9 元,而最优只有 550.2 元。

数据口径:书中参数为示意;全部计算由书中程序生成;本书不是法律、财务或经营上的个人化建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(18 段程序)、原典与作者的假说。

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: 1232.9, 550.2

AUC = 0.856 (d' = 1.5)
margint*to humanbad caughtgood sentcost/100at t=0.75
400.290.2130.6130.113195.9217.6
1501.170.4720.8790.370382.4417.0
6002.090.7750.9820.723550.21232.9
24003.020.9480.9990.935623.14496.3
the threshold moves with the margin; one global cutoff cannot
be optimal for both a 40-yuan and a 2400-yuan order
查看完整验算程序
# Chapter 4. p41: a cost-sensitive handoff threshold on the ROC curve.
# Model (illustrative parameters): the AI reports a confidence score s.
# On chats it can carry ("good") s ~ N(DP, 1); on chats it will fumble
# ("bad") s ~ N(0, 1); a share P_BAD are bad.  Rule: hand to a human
# when s < t.  A missed bad chat costs LOSS x margin (lost order plus
# complaint risk); a needless handoff costs C yuan of human time.
# The optimum sets the ROC slope to P(good) C / (P(bad) LOSS V), so
#   t* = DP / 2 - ln(P(good) C / (P(bad) LOSS V)) / DP.
# The fixed rule "t = DP / 2" (best plain accuracy at equal shares) is
# shown for comparison.  Costs are per 100 chats.
import math
from statistics import NormalDist

ND = NormalDist()
DP, P_BAD, LOSS, C = 1.5, 0.2, 0.4, 8.0
P_GOOD = 1.0 - P_BAD


def cost(t, v):
    miss = P_BAD * (1.0 - ND.cdf(t)) * LOSS * v
    waste = P_GOOD * ND.cdf(t - DP) * C
    return 100.0 * (miss + waste)


print("AUC = {:.3f} (d' = {})".format(ND.cdf(DP / math.sqrt(2.0)), DP))
print(" margin   t*    to human  bad caught  good sent  cost/100"
      "  at t=0.75")
for v in (40, 150, 600, 2400):
    k = P_GOOD * C / (P_BAD * LOSS * v)
    t = DP / 2.0 - math.log(k) / DP
    red = P_BAD * ND.cdf(t) + P_GOOD * ND.cdf(t - DP)
    print("{:7d}  {:5.2f}  {:8.3f}  {:10.3f}  {:9.3f}  {:8.1f}  {:9.1f}"
          .format(v, t, red, ND.cdf(t), ND.cdf(t - DP), cost(t, v),
                  cost(0.75, v)))
print("the threshold moves with the margin; one global cutoff cannot")
print("be optimal for both a 40-yuan and a 2400-yuan order")

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《私域对话队列耐心衰减分布、人机协同红黄绿灯移交门限与客诉升级马尔可夫阻断》第 4 章「红黄绿灯:ROC 曲线上的代价敏感移交门限」(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-175 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026175,
  author    = {Fan, Yafu},
  title     = {Patience-Decay Distributions in Private-Message Queues, Red-Amber-Green Human-AI Handoff Thresholds and Markov Blocking of Complaint Escalation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论