关于「实证闭环:店铺排序算法、求评时机与证伪方案」,《星级自选择与贝叶斯排序》给出的定量结论是什么?
结论:第八章给出店铺流程:向全部买家发同一封邀请、只邀满意买家会使均分虚高 +0.583 的红线、后验下界排序、蒸馏小工具「评论排序器」,以及每组 53,211 次访问的预先登记检验和推翻本书的条件。书中以标注框区分已发表的证据、可运行的 Python 计算(21 段程序)、原典与作者的假说。
数据口径:书中的顾客、商品与参数均为示意模型;全部计算由书中程序生成;本书是非个人化的经营参考,不是针对具体店铺的经营建议,也不是法律意见,书中提到的法规以上架前人工复核为准;本书反对一切刷单、购买评论与筛选式求评。 第一章用「购买偏差加少报偏差」的两段筛子解释评论星级的 J 型分布,示意参数下购买偏差使均值上移 +0.532 星。
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: 0.583, 21
| 90-day honest mean 4.064; mean at delivery 4.330 | ||||
| day | reviews | mean | bias | faults in |
| 1 | 288.2 | 4.309 | +0.245 | 0.080 |
| 3 | 266.1 | 4.271 | +0.207 | 0.221 |
| 7 | 226.7 | 4.212 | +0.149 | 0.442 |
| 14 | 171.4 | 4.147 | +0.083 | 0.689 |
| 21 | 129.5 | 4.110 | +0.046 | 0.826 |
| 30 | 90.4 | 4.085 | +0.022 | 0.918 |
| 45 | 49.6 | 4.070 | +0.006 | 0.976 |
| first day with |bias| < 0.05: 21 | ||||
| gated mean exceeds the 90-day target by +0.583 |
查看完整验算程序
# Chapter 8. p81: when to ask every buyer for a review, and why not to gate.
# Per 1000 orders (illustrative): at delivery the star mix is BUY0; a unit
# develops a fault with prob 0.08, at a time exponential with mean 12
# days, after which that buyer's star is 1. One request goes to every
# buyer on day d; the reply rate is 0.30 exp(-d / 25). The 90-day mean
# is the honest target. "Gated" = asking only buyers who first said they
# were happy (4 or 5 stars), a practice this book treats as a red line.
import math
BUY0 = (0.02, 0.03, 0.10, 0.30, 0.55)
FAULT, TAU, R0, DECAY = 0.08, 12.0, 0.30, 25.0
def mix_on(day):
f = FAULT * (1.0 - math.exp(-day / TAU))
sh = [(1.0 - f) * b for b in BUY0]
sh[0] += f
return sh
def mean(sh):
return math.fsum((k + 1) * s for k, s in enumerate(sh))
target = mean(mix_on(90))
print("90-day honest mean %.3f; mean at delivery %.3f"
% (target, mean(BUY0)))
print("%5s %9s %8s %8s %10s" % ("day", "reviews", "mean", "bias",
"faults in"))
for d in (1, 3, 7, 14, 21, 30, 45):
sh = mix_on(d)
rate = R0 * math.exp(-d / DECAY)
seen = 1.0 - math.exp(-d / TAU)
print("%5d %9.1f %8.3f %+8.3f %10.3f" % (
d, 1000 * rate, mean(sh), mean(sh) - target, seen))
loss = [(abs(mean(mix_on(d)) - target), d) for d in range(1, 61)]
print("first day with |bias| < 0.05: %d"
% min(d for b, d in loss if b < 0.05))
print("gated mean exceeds the 90-day target by %+.3f"
% (mean([0, 0, 0] + list(mix_on(90)[3:])) /
math.fsum(mix_on(90)[3:]) - target))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《评论星级J型自选择偏差、贝叶斯平均排序与九品中正评级通胀母典》第 8 章「实证闭环:店铺排序算法、求评时机与证伪方案」(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-212 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026212,
author = {Fan, Yafu},
title = {J-Shaped Self-Selection in Star Ratings, Bayesian-Average Ranking and Rating Inflation from the Nine-Rank System to Online Marketplaces},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}