关于「疲劳曲线不是一条指数线:重复曝光的两因素过程与半衰期估计」,《素材疲劳与正交混剪》给出的定量结论是什么?
结论:第一章把三种疲劳曲线写成公式,从合成日志拟合半衰期与自助法区间,并用新观众校正环境,示意参数下环境漂移占汇总下滑的 0.211。
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: 0.211
| naive daily half-life of aggregate CTR | 6.95 days |
| env drift from fresh viewers | -0.0140 per day (true -0.0200) |
| env-adjusted frequency half-life | 3.93 exposures (true 4.00) |
| log drop day 1 to 14 | -1.244; environment share 0.211 |
查看完整验算程序
# Chapter 1. p13: audience fatigue versus a drifting auction environment
# Daily CTR = env(day) * g(frequency). Cohorts of new viewers enter each
# day and see one impression a day while they stay. The environment
# multiplier env falls 2 per cent a day (competition, season). A naive
# reader fits a half-life to the daily aggregate CTR; a careful reader
# uses frequency-1 impressions (fresh viewers) to track env first.
import math
import random
rng = random.Random(40013)
DAYS = 14
H_TRUE = 4.0
def g(f):
return 0.5 ** ((f - 1) / H_TRUE)
def env(d):
return 0.98 ** (d - 1)
def binom(n, p):
k = rng.gauss(n * p, math.sqrt(n * p * (1.0 - p)))
return max(0, round(k))
new = [round(60000 * 0.8 ** (d - 1)) for d in range(1, DAYS + 1)]
daily = [] # (day, impressions, clicks, fresh_n, fresh_clicks)
by_f = {} # frequency -> list of (impr, clicks, day)
for d in range(1, DAYS + 1):
imps, clk = [], []
for j in range(1, d + 1):
f = d - j + 1
n = round(new[j - 1] * 0.85 ** (f - 1)) # some viewers leave
k = binom(n, 0.012 * env(d) * g(f))
imps.append(n)
clk.append(k)
by_f.setdefault(f, []).append((n, k, d))
daily.append((d, math.fsum(imps), math.fsum(clk), imps[-1], clk[-1]))
def slope(xs, ys, ws):
sw = math.fsum(ws)
mx = math.fsum(w * x for w, x in zip(ws, xs)) / sw
my = math.fsum(w * y for w, y in zip(ws, ys)) / sw
sxy = math.fsum(w * (x - mx) * (y - my) for w, x, y in zip(ws, xs, ys))
sxx = math.fsum(w * (x - mx) ** 2 for w, x in zip(ws, xs))
return sxy / sxx
xs = [r[0] for r in daily]
ys = [math.log(r[2] / r[1]) for r in daily]
b_naive = slope(xs, ys, [r[2] for r in daily])
print("naive daily half-life of aggregate CTR: {:.2f} days".format(
math.log(0.5) / b_naive))
env_hat = {r[0]: (r[4] / r[3]) for r in daily}
b_env = slope(xs, [math.log(env_hat[d]) for d in xs], [r[4] for r in daily])
print("env drift from fresh viewers: {:.4f} per day (true {:.4f})".format(
math.exp(b_env) - 1.0, 0.98 - 1.0))
fx, fy, fw = [], [], []
for f in sorted(by_f):
for n, k, d in by_f[f]:
if k > 0:
fx.append(f)
fy.append(math.log(k / n / env_hat[d]))
fw.append(k)
b_f = slope(fx, fy, fw)
print("env-adjusted frequency half-life: {:.2f} exposures (true {:.2f})"
.format(math.log(0.5) / b_f, H_TRUE))
first, last = daily[0], daily[-1]
drop = math.log((last[2] / last[1]) / (first[2] / first[1]))
env_part = math.log(env(DAYS) / env(1))
print("log drop day 1 to {}: {:.3f}; environment share {:.3f}".format(
DAYS, drop, env_part / drop))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 1 章「疲劳曲线不是一条指数线:重复曝光的两因素过程与半衰期估计」(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}
}