关于「评分通胀:规范追逐、信息衰减与求评摩擦」,《星级自选择与贝叶斯排序》给出的定量结论是什么?
结论:第六章用规范追逐模型说明评分通胀:诚实感受不变,均分九年间从 3.485 升到 4.090,分辨两个卖家所需的评论随之增加;并讨论求评摩擦如何决定谁开口。
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: 3.485, 4.090
| norm cost c = 0.5, year by year | |||||
| year | mean | sd | 5-star | >=4 | corr |
| 0 | 3.485 | 0.914 | 0.133 | 0.499 | 0.944 |
| 1 | 3.575 | 0.907 | 0.157 | 0.541 | 0.942 |
| 2 | 3.660 | 0.899 | 0.182 | 0.580 | 0.940 |
| 3 | 3.740 | 0.889 | 0.208 | 0.617 | 0.936 |
| 4 | 3.815 | 0.879 | 0.235 | 0.651 | 0.933 |
| 5 | 3.884 | 0.868 | 0.262 | 0.682 | 0.929 |
| 6 | 3.946 | 0.857 | 0.288 | 0.709 | 0.925 |
| 7 | 4.000 | 0.846 | 0.312 | 0.732 | 0.921 |
| 8 | 4.048 | 0.835 | 0.334 | 0.752 | 0.917 |
| 9 | 4.090 | 0.825 | 0.354 | 0.770 | 0.913 |
| state in year 10 by norm cost c | |||||
| c | mean | sd | 5-star | >=4 | corr |
| 0.00 | 3.309 | 0.924 | 0.094 | 0.419 | 0.947 |
| 0.25 | 3.786 | 0.883 | 0.224 | 0.638 | 0.934 |
| 0.50 | 4.090 | 0.825 | 0.354 | 0.770 | 0.913 |
| 1.00 | 4.401 | 0.722 | 0.530 | 0.882 | 0.866 |
查看完整验算程序
# Chapter 6. p61: a norm-chasing rater inflates the scale year after year.
# Honest assessment x ~ N(3.4, 0.9) (illustrative). Last year raters on
# average gave L stars more than the honest view, so the going rate for
# an experience x is x + L, and anything below it reads as a complaint.
# A rater picks r in 1..5 to minimise
# (r - x - L)^2 + c * max(0, x + L - r) / S:
# distance from the going rate plus the harm of a below-norm rating, in
# units of last year's spread S of displayed ratings (after Filippas,
# Horton and Golden: harsh feedback grows costlier as others' rises).
# Each year L and S are updated from the ratings just given. Exact
# integration over x on a midpoint grid; no randomness.
import math
MU, SD = 3.4, 0.9
H = 0.002
GRID = [MU - 6 * SD + (i + 0.5) * H for i in range(int(12 * SD / H))]
DENS = [math.exp(-0.5 * ((x - MU) / SD) ** 2) / (SD * math.sqrt(2 * math.pi))
* H for x in GRID]
def choose(x, c, lean, spread):
best, val = 1, None
for r in range(1, 6):
v = (r - x - lean) ** 2 + c * max(0.0, x + lean - r) / spread
if val is None or v < val:
best, val = r, v
return best
def year(c, lean, spread):
rs = [choose(x, c, lean, spread) for x in GRID]
m = math.fsum(d * r for d, r in zip(DENS, rs))
s5 = math.fsum(d for d, r in zip(DENS, rs) if r == 5)
s4 = math.fsum(d for d, r in zip(DENS, rs) if r >= 4)
cov = math.fsum(d * (x - MU) * (r - m)
for d, x, r in zip(DENS, GRID, rs))
vr = math.fsum(d * (r - m) ** 2 for d, r in zip(DENS, rs))
return m, s5, s4, cov / (SD * math.sqrt(vr)), math.sqrt(vr), m - MU
def history(c, years):
_, _, _, _, spread, lean = year(0.0, 0.0, 1.0)
out = []
for _ in range(years):
m, s5, s4, rho, spread, lean = year(c, lean, spread)
out.append((m, spread, s5, s4, rho))
return out
print("norm cost c = 0.5, year by year")
print("%4s %7s %7s %7s %7s %8s" % ("year", "mean", "sd", "5-star", ">=4",
"corr"))
for t, row in enumerate(history(0.5, 10)):
print("%4d %7.3f %7.3f %7.3f %7.3f %8.3f" % ((t,) + row))
print("state in year 10 by norm cost c")
print("%5s %7s %7s %7s %7s %8s" % ("c", "mean", "sd", "5-star", ">=4",
"corr"))
for c in (0.0, 0.25, 0.5, 1.0):
print("%5.2f %7.3f %7.3f %7.3f %7.3f %8.3f" % ((c,) + history(c, 10)[-1]))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 6 章「评分通胀:规范追逐、信息衰减与求评摩擦」(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}
}