关于「正交表能估什么:L4、L8、L9、L16 的构造、主效应效率与混叠结构」,《素材疲劳与正交混剪》给出的定量结论是什么?
结论:第三章从零构造 L4、L8、L9、L16,L9 的对比方差为 0.6667,逐个改因素为 2.0000,但钩子与召唤的配合会全部落到痛点模块上。
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。
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: 0.6667, 2.0000
contrast variance, level 1 minus level 0, noise sd 1: L9 : 0.6667 (3 runs per level mean) OFAT: 2.0000 (one run against the base run) Monte Carlo 20,000 runs, A1-A0 (true 0.3): L9 mean 0.2925 variance 0.6706 OFAT mean 0.3025 variance 1.9962 synergy bonus 0.2 when hook level = CTA level; level means: hook L9 0.067 0.067 0.067 full 0.067 0.067 0.067 pain L9 0.200 0.000 0.000 full 0.067 0.067 0.067 proof L9 0.067 0.067 0.067 full 0.067 0.067 0.067 cta L9 0.067 0.067 0.067 full 0.067 0.067 0.067
查看完整验算程序
# Chapter 3. p33: what L9 buys (precision) and what it costs (aliasing)
# Part 1: variance of a level contrast with nine runs, L9 versus
# one-factor-at-a-time (OFAT: a base run plus two runs per factor).
# Part 2: a hook-by-CTA synergy (bonus when hook level equals CTA level)
# in L9 columns A=i, B=j, C=i+j, D=i+2j (mod 3) versus the full 81.
import itertools
import math
import random
SIG = 1.0
print("contrast variance, level 1 minus level 0, noise sd 1:")
print(" L9 : {:.4f} (3 runs per level mean)".format(SIG * (1 / 3 + 1 / 3)))
print(" OFAT: {:.4f} (one run against the base run)".format(SIG * 2))
rng = random.Random(40033)
TRUE = [0.0, 0.3, 0.1] # factor A level effects, others zero
l9 = [[i, j, (i + j) % 3, (i + 2 * j) % 3]
for i, j in itertools.product(range(3), repeat=2)]
ofat = [[0, 0, 0, 0]]
for f in range(4):
for lv in (1, 2):
r = [0, 0, 0, 0]
r[f] = lv
ofat.append(r)
e9, eo = [], []
for _ in range(20000):
y9 = [TRUE[r[0]] + rng.gauss(0, SIG) for r in l9]
m1 = math.fsum(y for y, r in zip(y9, l9) if r[0] == 1) / 3
m0 = math.fsum(y for y, r in zip(y9, l9) if r[0] == 0) / 3
e9.append(m1 - m0)
yo = [TRUE[r[0]] + rng.gauss(0, SIG) for r in ofat]
eo.append(yo[1] - yo[0])
print("Monte Carlo 20,000 runs, A1-A0 (true 0.3):")
for name, es in (("L9 ", e9), ("OFAT", eo)):
m = math.fsum(es) / len(es)
v = math.fsum((e - m) ** 2 for e in es) / (len(es) - 1)
print(" {} mean {:.4f} variance {:.4f}".format(name, m, v))
BONUS = 0.2
full = list(itertools.product(range(3), repeat=4))
def level_means(rows, f):
out = []
for lv in range(3):
ys = [BONUS * (r[0] == r[3]) for r in rows if r[f] == lv]
out.append(math.fsum(ys) / len(ys))
return out
print("synergy bonus {:.1f} when hook level = CTA level; level means:"
.format(BONUS))
for f, name in enumerate(["hook", "pain", "proof", "cta"]):
a = level_means(l9, f)
b = level_means(full, f)
print(" {:<5} L9 {} full {}".format(
name, " ".join("{:.3f}".format(x) for x in a),
" ".join("{:.3f}".format(x) for x in b)))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 3 章「正交表能估什么:L4、L8、L9、L16 的构造、主效应效率与混叠结构」(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}
}