关于「平台不会替你做随机实验:按预测分配曝光的偏差与汤普森抽样赛马」,《素材疲劳与正交混剪》给出的定量结论是什么?

结论:第六章说明平台自适应分配下,两条相同素材报表上的相对差异平均偏离真值 0.348;汤普森抽样把遗憾从均分的 317.8 次点击降到中位数 187.1 次,代价是差异估计有偏。

数据口径:买量圈流行的说法是:扩量期素材衰退得快,因为观众重复曝光使点击率按可估计的半衰期衰减;把短视频拆成开头钩子、痛点、证据、行动召唤四个模块,用正交表组合,十几个镜头就能自动合成上百条低相关的测试素材,突破产能瓶颈。 书中参数为示意,全部计算由书中程序生成。

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: 187.1, 317.8

impressions per campaign 200,000
equal split regret 317.8 clicks
Thompson regret 5/50/95%: 118.8 187.1 301.3 clicks
Thompson: best creative most shown in 0.793 of campaigns
best creative impression share 5/50/95%: 0.070 0.435 0.665
查看完整验算程序
# Chapter 6. p62: Thompson sampling versus an equal split, nine creatives
# Nine L9 creatives; true CTRs are illustrative (one at 0.0120, the rest
# between 0.0095 and 0.0110). 200 rounds of 1,000 impressions. Batched
# Thompson sampling (Thompson 1933): each round draw one Beta posterior
# sample per creative and give the whole batch to the largest draw.
# Regret = clicks lost against always showing the best creative.
import math
import random

rng = random.Random(40062)
TRUE = [0.0120, 0.0110, 0.0108, 0.0105, 0.0103, 0.0100, 0.0099, 0.0097,
        0.0095]
ROUNDS, BATCH, RUNS = 200, 1000, 300
BEST = max(TRUE)


def clicks(n, p):
    return max(0, round(rng.gauss(n * p, math.sqrt(n * p * (1 - p)))))


def thompson():
    a = [1] * 9
    b = [1] * 9
    regret = []
    for _ in range(ROUNDS):
        draws = [rng.betavariate(a[k], b[k]) for k in range(9)]
        k = draws.index(max(draws))
        c = clicks(BATCH, TRUE[k])
        a[k] += c
        b[k] += BATCH - c
        regret.append((BEST - TRUE[k]) * BATCH)
    pulls = [a[k] + b[k] - 2 for k in range(9)]
    return math.fsum(regret), pulls.index(max(pulls)) == 0, pulls[0]


def equal_split():
    per = ROUNDS * BATCH / 9
    return math.fsum((BEST - p) * per for p in TRUE)


res = [thompson() for _ in range(RUNS)]
reg = sorted(r[0] for r in res)
share = sorted(r[2] / (ROUNDS * BATCH) for r in res)
print("impressions per campaign {:,}".format(ROUNDS * BATCH))
print("equal split regret {:.1f} clicks".format(equal_split()))
print("Thompson regret 5/50/95%: {:.1f} {:.1f} {:.1f} clicks".format(
    reg[RUNS // 20], reg[RUNS // 2], reg[RUNS * 19 // 20]))
print("Thompson: best creative most shown in {:.3f} of campaigns".format(
    len([1 for r in res if r[1]]) / RUNS))
print("best creative impression share 5/50/95%: {:.3f} {:.3f} {:.3f}".format(
    share[RUNS // 20], share[RUNS // 2], share[RUNS * 19 // 20]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《投放素材频次疲劳半衰期、分镜基因正交重组与百条广告全自动混剪母典》第 6 章「平台不会替你做随机实验:按预测分配曝光的偏差与汤普森抽样赛马」(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-400 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026400,
  author    = {Fan, Yafu},
  title     = {Ad-Creative Fatigue Half-Lives, Orthogonal Storyboard Recombination and Automated Permutation Rendering},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论