关于「注意力生存曲线:删失、风险函数与前三秒的杠杆」,《直播开场、评论区与收口的计算》给出的定量结论是什么?

结论:第二章把留存曲线还原为有删失的生存函数,前三秒的同幅改动对平均观看时长的增益是后段的 3.65 倍。

数据口径:书中参数均为示意;全部计算由书中程序生成;本书不构成个别化的法律、广告合规或营销建议。 第一章用伽马-泊松模型计算开场七个音节的贝叶斯惊奇度,留存与惊奇度呈倒 U 关系,示意模型的最优值为 1.79 比特,九成增益区间为 1.13 到 2.62 比特,「七个字」取决于语速。

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: 3.65

interval (s) hazard per s share of swipes
0 to1 0.240 0.299
1 to2 0.184 0.181
2 to3 0.117 0.100
3 to5 0.081 0.117
5 to 100.035 0.103
10 to 200.029 0.128
20 to 300.025 0.071
scenario completion mean watch (s)
as estimated0.242 11.92
first 3 s -20 %0.269 13.16
10 to 30 s -20 %0.269 12.26
front edit / back edit, gain in mean watch3.65
查看完整验算程序
# Chapter 2. p22: Where the swipe hazard lives, and where to spend effort.
# Same 400 simulated viewers as p21 (same seed, same mixture, same 30 %
# interruptions). Piecewise-constant hazard on seven intervals is the
# maximum-likelihood estimate: swipes / person-seconds at risk.
# Then two edits of equal size on paper: cut the hazard by 20 % in the
# first 3 s, or cut it by 20 % from 10 s to 30 s. The survival curve
# implied by the estimated hazards gives completion and mean watch time.
# All inputs are illustrative.
import math
import random

rng = random.Random(20261112)
N, END = 400, 30.0


def draw():
    if rng.random() < 0.45:
        t = rng.expovariate(1 / 1.8)
    else:
        t = rng.weibullvariate(40.0, 0.9)
    if rng.random() < 0.30:
        cut = rng.uniform(0.0, END)
        if cut < t:
            return cut, 0
    if t >= END:
        return END, 0
    return t, 1


data = [draw() for _ in range(N)]
CUTS = [0, 1, 2, 3, 5, 10, 20, 30]
haz = []
for lo, hi in zip(CUTS, CUTS[1:]):
    expo = math.fsum(min(max(t - lo, 0.0), hi - lo) for t, _ in data)
    ev = len([1 for t, e in data if e and lo <= t < hi])
    haz.append(ev / expo)


def curve(h):
    s, parts = 1.0, []
    for (lo, hi), hz in zip(zip(CUTS, CUTS[1:]), h):
        w = hi - lo
        parts.append(s * (1 - math.exp(-hz * w)) / hz)
        s *= math.exp(-hz * w)
    return s, math.fsum(parts)


print("interval (s)   hazard per s   share of swipes")
tot = len([1 for _, e in data if e])
for (lo, hi), hz in zip(zip(CUTS, CUTS[1:]), haz):
    ev = len([1 for t, e in data if e and lo <= t < hi])
    print("  {:2d} to {:2d}       {:8.3f}        {:.3f}".format(
        lo, hi, hz, ev / tot))
base = curve(haz)
front = curve([h * 0.8 if i < 3 else h for i, h in enumerate(haz)])
back = curve([h * 0.8 if i >= 5 else h for i, h in enumerate(haz)])
print("scenario            completion   mean watch (s)")
for name, (c, m) in (("as estimated", base), ("first 3 s -20 %", front),
                     ("10 to 30 s -20 %", back)):
    print("  {:<17} {:9.3f}   {:12.2f}".format(name, c, m))
print("front edit / back edit, gain in mean watch: {:.2f}".format(
    (front[1] - base[1]) / (back[1] - base[1])))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《短视频与直播前七字听觉劫持、评论区从众引爆与脉冲逼单剧本学》第 2 章「注意力生存曲线:删失、风险函数与前三秒的杠杆」(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-147 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026147,
  author    = {Fan, Yafu},
  title     = {Seven-Syllable Auditory Hooks in Short Video and Livestreams, Comment-Section Herding Ignition and Pulse-Closing Scripts},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论