关于「停投会恢复:纳伊克型动态、频次上限与轮换条数」,《素材疲劳与正交混剪》给出的定量结论是什么?

结论:第二章用纳伊克型模型估出磨损半衰期 8.42 个在播日、恢复半程 3.92 个停播日,并推出频次上限与轮换条数。

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: 3.92, 8.42

true a 0.080 r 0.150; fitted a 0.079 r 0.162 c 0.01172
bootstrap 90%: a 0.074 to 0.084; r 0.144 to 0.178
wear-out half-life 8.42 on-air days
recovery half-time 3.92 rest days
30 on-air days: continuous mean q 0.383, pulsed mean q 0.616
查看完整验算程序
# Chapter 2. p21: a Naik-type quality path with wear-out and recovery
# Daily quality q: on air  q' = q - a * q        (wear-out)
#                  off air q' = q + r * (1 - q)  (restoration)
# CTR = c * q. After Naik, Mantrala and Sawyer (1998); parameters are
# illustrative. Fit (a, r) by a coarse-then-fine grid on the Poisson
# likelihood with c profiled out; interval by parametric bootstrap.
import math
import random

rng = random.Random(40021)
SCHED = [1] * 10 + [0] * 7 + [1] * 10 + [0] * 7 + [1] * 10
NDAY = 50000
A_T, R_T, C_T = 0.08, 0.15, 0.012


def path(a, r, sched):
    q, out = 1.0, []
    for u in sched:
        out.append(q)
        q = q - a * q if u else q + r * (1.0 - q)
    return out


def sim(a, r, c):
    qs = path(a, r, SCHED)
    ks = []
    for u, q in zip(SCHED, qs):
        if u:
            m = NDAY * c * q
            ks.append(max(0, round(rng.gauss(m, math.sqrt(m)))))
        else:
            ks.append(0)
    return ks


def loglik(a, r, ks):
    qs = path(a, r, SCHED)
    on = [(k, q) for u, k, q in zip(SCHED, ks, qs) if u]
    c = math.fsum(k for k, _ in on) / math.fsum(NDAY * q for _, q in on)
    return math.fsum(k * math.log(c * q) - NDAY * c * q for k, q in on), c


def fit(ks):
    best = max(((loglik(0.01 * i, 0.02 * j, ks)[0], 0.01 * i, 0.02 * j)
                for i in range(1, 21) for j in range(1, 26)))
    _, a0, r0 = best
    fine = max(((loglik(a0 + 0.001 * i, r0 + 0.002 * j, ks)[0],
                 a0 + 0.001 * i, r0 + 0.002 * j)
                for i in range(-10, 11) for j in range(-10, 11)
                if a0 + 0.001 * i > 0 and r0 + 0.002 * j > 0))
    return fine[1], fine[2]


ks = sim(A_T, R_T, C_T)
a_hat, r_hat = fit(ks)
c_hat = loglik(a_hat, r_hat, ks)[1]
print("true a {:.3f} r {:.3f}; fitted a {:.3f} r {:.3f} c {:.5f}".format(
    A_T, R_T, a_hat, r_hat, c_hat))
ba, br = [], []
for _ in range(100):
    a_b, r_b = fit(sim(a_hat, r_hat, c_hat))
    ba.append(a_b)
    br.append(r_b)
ba.sort()
br.sort()
print("bootstrap 90%: a {:.3f} to {:.3f}; r {:.3f} to {:.3f}".format(
    ba[4], ba[94], br[4], br[94]))
print("wear-out half-life {:.2f} on-air days".format(
    math.log(0.5) / math.log(1.0 - a_hat)))
print("recovery half-time {:.2f} rest days".format(
    math.log(0.5) / math.log(1.0 - r_hat)))
cont = path(A_T, R_T, [1] * 30)
puls = [q for u, q in zip(SCHED, path(A_T, R_T, SCHED)) if u]
print("30 on-air days: continuous mean q {:.3f}, pulsed mean q {:.3f}"
      .format(math.fsum(cont) / 30, math.fsum(puls) / 30))

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 2 章「停投会恢复:纳伊克型动态、频次上限与轮换条数」(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