关于「长尾问题词:Zipf 需求、选题打分与「单支小、合计大」的组合风险」,《长视频半衰期与深度信任》给出的定量结论是什么?

结论:第三章用 Zipf 分布与选题打分处理长尾问题词,十支冷门问题视频胜过两支大词视频的概率是 0.722。

数据口径:示意曲线下,短视频头 48 小时拿到三年观看的 0.750,带搜索地板的长视频只拿到 0.059,第 1,095 天仍有每天 5.00 次观看;半衰期翻十倍时,短内容峰值要高出 9.96 倍才打平现值。 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: 0.722

decay factor (sum of 36 monthly weights)15.14
36-month views mean P10 P50 P90
2 head6,948 1,817 1,817 19,075
10 tail3,889 2,652 3,760 5,289
P(10 tail > 2 head): views 0.722 weighted 0.725
coefficient of variation: head 1.246 tail 0.268
查看完整验算程序
# Chapter 3. p33: two head-term videos or ten long-tail videos, 60 hours.
# Head term: 20,000 searches a month; a video lands on the first screen
# with probability 0.15 and then captures 0.06 of searches, otherwise
# 0.003. Each head video costs 30 hours.
# Tail term: Zipf rank r uniform on 50..500, searches 20,000 / r a month;
# capture uniform on 0.10..0.40 (thin competition). Each costs 6 hours.
# Monthly views decay with a 365-day half-life; 36 months are summed.
# w = qualified actions per view, 1.0 for head terms and 2.5 for tail
# questions (closer to a decision). 20,000 Monte Carlo runs, fixed seed.
# All values are illustrative.
import math
import random

RNG = random.Random(20261120)
DECAY = math.fsum(2.0 ** (-(m + 0.5) * 30.4375 / 365.0) for m in range(36))


def head():
    share = 0.06 if RNG.random() < 0.15 else 0.003
    return 20000.0 * share * DECAY


def tail():
    r = RNG.randint(50, 500)
    return 20000.0 / r * RNG.uniform(0.10, 0.40) * DECAY


def q(xs, p):
    return xs[int(p * (len(xs) - 1))]


if __name__ == "__main__":
    H, T, WH, WT = [], [], [], []
    wins_v = wins_w = 0
    for _ in range(20000):
        h = math.fsum([head(), head()])
        t = math.fsum([tail() for _ in range(10)])
        H.append(h)
        T.append(t)
        WH.append(h * 1.0)
        WT.append(t * 2.5)
        wins_v += t > h
        wins_w += t * 2.5 > h
    print("decay factor (sum of 36 monthly weights): {:.2f}".format(DECAY))
    print("36-month views     mean      P10      P50      P90")
    for name, xs in (("2 head", H), ("10 tail", T)):
        mean = math.fsum(xs) / len(xs)
        xs = sorted(xs)
        print("  {:<8}  {:>9,.0f} {:>8,.0f} {:>8,.0f} {:>8,.0f}".format(
            name, mean, q(xs, 0.1), q(xs, 0.5), q(xs, 0.9)))
    print("P(10 tail > 2 head): views {:.3f}  weighted {:.3f}".format(
        wins_v / 20000.0, wins_w / 20000.0))
    sd = {}
    for name, xs in (("head", H), ("tail", T)):
        m = math.fsum(xs) / len(xs)
        sd[name] = math.sqrt(math.fsum((x - m) ** 2 for x in xs)
                             / (len(xs) - 1)) / m
    print("coefficient of variation: head {:.3f}  tail {:.3f}".format(
        sd["head"], sd["tail"]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《长视频搜索半衰期红利、深度信任预售与多年期被动获客资产母典》第 3 章「长尾问题词:Zipf 需求、选题打分与「单支小、合计大」的组合风险」(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-370 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026370,
  author    = {Fan, Yafu},
  title     = {Long-Form Video Search Half-Lives, Deep-Trust Pre-Selling and Multi-Year Passive Acquisition Assets},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论