5D Library › 短视频完播率风险函数、推荐反馈环与自动化获客漏斗 · English
第三章 反馈环、马太效应与可被「刷」的指标
推荐反馈环中的马太效应
推荐系统把曝光给已经获得点赞的视频,点赞又带来更多曝光。这是一个自我强化的反馈环。Merton 用「马太效应」描述科学界中已有声望者获得更多承认的现象 (Merton 1968);在网络科学中,「富者愈富」的优先连接会产生幂律分布 (Barabási and Albert 1999);最简单的数学模型是波利亚罐子 (Eggenberger and Pólya 1923)。Salganik、Dodds 与 Watts 用一个音乐市场实验证明:让听众看到他人的下载数,会使成功更不平等,也更难预测 (Salganik et al. 2006)。
程序设 48 条视频的吸引力从 0.2 均匀分布到 0.8,10,000 名观众每人被推送一条视频,被推送的概率与(点赞数 + 1)的 次方成正比; 是没有反馈的世界, 是波利亚罐子, 让人气复利增长。每种 各运行八个独立的「世界」。
推导
# Chapter 3. p05: The Matthew effect in a recommender feedback loop.
# 48 videos with fixed appeal q_i spread evenly from 0.2 to 0.8. Each
# of 10,000 viewers is shown one video, chosen with probability
# proportional to (likes_i + 1)^b, and likes it with chance q_i.
# b = 0 is a world with no feedback (independent exposure); b = 1 is a
# Polya urn; b = 1.5 lets popularity compound. For each b we run eight
# independent worlds with the same videos (as Salganik et al. did with
# songs) and report: the Gini coefficient of likes, the share of likes
# held by the top video, the rank correlation of likes with appeal,
# and in how many of the eight worlds the best video came first.
import random
def gini(xs):
xs = sorted(xs)
n = len(xs)
cum = sum((i + 1) * x for i, x in enumerate(xs))
return (2 * cum) / (n * sum(xs)) - (n + 1) / n
def ranks(xs):
order = sorted(range(len(xs)), key=lambda i: xs[i])
r = [0.0] * len(xs)
for pos, i in enumerate(order):
r[i] = float(pos)
return r
def spearman(a, b):
ra, rb = ranks(a), ranks(b)
n = len(a)
ma, mb = sum(ra) / n, sum(rb) / n
cov = sum((x - ma) * (y - mb) for x, y in zip(ra, rb))
va = sum((x - ma) ** 2 for x in ra)
vb = sum((y - mb) ** 2 for y in rb)
return cov / (va * vb) ** 0.5
def world(b, rng, n=48, viewers=10000):
q = [0.2 + 0.6 * i / (n - 1) for i in range(n)]
likes = [0] * n
for _ in range(viewers):
w = [(x + 1) ** b for x in likes]
i = rng.choices(range(n), weights=w)[0]
if rng.random() < q[i]:
likes[i] += 1
return q, likes
if __name__ == "__main__":
rng = random.Random(1968)
print(" b Gini top share rank corr best video first")
for b in (0.0, 1.0, 1.5):
g = t = c = 0.0
wins = 0
for _ in range(8):
q, likes = world(b, rng)
g += gini(likes) / 8
t += max(likes) / sum(likes) / 8
c += spearman(q, likes) / 8
wins += likes.index(max(likes)) == len(q) - 1
print(" {:3.1f} {:8.3f} {:10.3f} {:11.3f} {:10d} of 8".format(
b, g, t, c, wins))
print("\nReading: without feedback, likes track appeal closely")
print("(rank correlation 0.970), and the best video wins in half the")
print("worlds only because its neighbours are almost as good. With")
print("feedback, early luck is amplified: inequality rises, the link")
print("between appeal and success weakens, and one video can take")
print("most of the likes. Past view counts measure luck as well as")
print("quality.")
# b Gini top share rank corr best video first
# 0.0 0.214 0.035 0.970 4 of 8
# 1.0 0.702 0.186 0.794 1 of 8
# 1.5 0.964 0.853 0.621 2 of 8
#
# Reading: without feedback, likes track appeal closely
# (rank correlation 0.970), and the best video wins in half the
# worlds only because its neighbours are almost as good. With
# feedback, early luck is amplified: inequality rises, the link
# between appeal and success weakens, and one video can take
# most of the likes. Past view counts measure luck as well as
# quality.没有反馈时,点赞的基尼系数为 0.214,点赞与吸引力的等级相关为 0.970;最佳视频只在一半的世界里夺冠,仅仅是因为它的邻居几乎一样好。 时基尼系数升到 0.702,等级相关降到 0.794; 时基尼系数达到 0.964,冠军独占 0.853 的点赞,等级相关只剩 0.621。在有反馈的世界里,过去的播放数既衡量质量,也衡量运气。
哪个指标能被剪辑长度刷高
Goodhart 指出,一个统计规律一旦被用作控制的目标,就会失去原有的规律性 (Goodhart 1975)。短视频的指标也一样。程序设五条长度与「留存力」(威布尔尺度 )各不相同的视频,比较四种排名:完播率、平均观看时长、按同时长基线( 秒)归一的完播率,以及由完播率直接反推的留存力。
推导
# Chapter 3. p06: Which metric can be gamed by video length?
# Five videos of different lengths D and different "holding power" L
# (the Weibull scale of p01, shape k = 0.6; larger L = viewers stay
# longer at every moment). The platform can rank them by:
# completion rate S(D);
# mean watch time (seconds);
# length-normalised completion S(D) / S0(D), where S0 is the
# platform-wide baseline curve with L = 20 s for that length;
# fitted holding power D / (-ln S(D))^(1/k), which inverts the curve.
# True quality here is L alone. A metric is gameable if a creator can
# raise it by changing length rather than holding power.
import math
K = 0.6
def surv(t, L):
return math.exp(-(t / L) ** K)
def mean_watch(D, L, n=2000):
return sum(surv((i + 0.5) * D / n, L) for i in range(n)) * D / n
VIDEOS = [("A", 7, 16.0), ("B", 15, 24.0), ("C", 30, 20.0),
("D", 60, 28.0), ("E", 90, 18.0)]
def order(vals):
return "".join(n for n, _ in sorted(vals, key=lambda x: -x[1]))
if __name__ == "__main__":
print(" video D (s) L (s) completion watch (s) normalised")
rows = []
for name, D, L in VIDEOS:
c, w, z = surv(D, L), mean_watch(D, L), surv(D, L) / surv(D, 20.0)
rows.append((name, L, c, w, z))
print(" {:5s} {:6d} {:7.1f} {:12.3f} {:11.1f} {:12.3f}".format(
name, D, L, c, w, z))
print("\n ranking by true holding power L : " +
order([(r[0], r[1]) for r in rows]))
print(" ranking by completion rate : " +
order([(r[0], r[2]) for r in rows]))
print(" ranking by mean watch time : " +
order([(r[0], r[3]) for r in rows]))
print(" ranking by normalised completion: " +
order([(r[0], r[4]) for r in rows]))
print(" ranking by fitted holding power : " + order(
[(n, D / (-math.log(surv(D, L))) ** (1 / K)) for n, D, L in VIDEOS]))
print("\n video A cut from 7 s to 5 s: completion {:.3f} -> {:.3f},"
.format(surv(7, 16.0), surv(5, 16.0)))
print(" normalised {:.3f} -> {:.3f}".format(
surv(7, 16.0) / surv(7, 20.0), surv(5, 16.0) / surv(5, 20.0)))
print("\nReading: completion rewards short videos and watch time")
print("rewards long ones, whatever their quality; each can be raised")
print("by editing length alone. Comparing each video with the")
print("baseline for its own length nearly recovers the true order")
print("(only the two weakest swap) and barely moves when a video is")
print("trimmed; fitting the holding power itself recovers it fully.")
# video D (s) L (s) completion watch (s) normalised
# A 7 16.0 0.544 4.8 0.927
# B 15 24.0 0.470 9.5 1.091
# C 30 20.0 0.279 14.3 1.000
# D 60 28.0 0.206 24.3 1.424
# E 90 18.0 0.072 22.0 0.851
#
# ranking by true holding power L : DBCEA
# ranking by completion rate : ABCDE
# ranking by mean watch time : DECBA
# ranking by normalised completion: DBCAE
# ranking by fitted holding power : DBCEA
#
# video A cut from 7 s to 5 s: completion 0.544 -> 0.608,
# normalised 0.927 -> 0.940
#
# Reading: completion rewards short videos and watch time
# rewards long ones, whatever their quality; each can be raised
# by editing length alone. Comparing each video with the
# baseline for its own length nearly recovers the true order
# (only the two weakest swap) and barely moves when a video is
# trimmed; fitting the holding power itself recovers it fully.真实的留存力次序是 D、B、C、E、A。完播率的排名恰好按长度从短到长(A、B、C、D、E):7 秒的 A 留存力最弱,完播率却最高,为 0.544。平均观看时长则偏爱长视频:90 秒的 E 留存力倒数第二,观看时长却排第二(22.0 秒)。按同时长基线归一的完播率几乎恢复了真实次序,只有最弱的两条互换;直接反推留存力则完全恢复。把 A 从 7 秒剪到 5 秒,完播率从 0.544 升到 0.608,归一指标只从 0.927 升到 0.940。
假说 · 尚未验证
可检验的推论。 若平台或团队以原始完播率考核创作者,创作者的视频时长分布应在考核开始后向短端移动,而按同时长归一的指标不应随之上升。什么会证明这一推论错:若时长分布不变,或归一指标与完播率同步上升,则完播率的上升来自内容改进而不是剪辑长度。
小结
推荐反馈把早期的运气放大为长期的差距,过去的播放数不能直接当作质量的证据。完播率与观看时长都能被剪辑长度操纵;比较视频时应按同时长基线归一,或直接拟合留存力。
参考文献
Fan, Y. (2026). Short-Video Completion Hazards, Recommender Exploration Loops and Automated Acquisition Funnels: Weibull Swipe-Away Kinetics, Multi-Stage Pool Promotion Thresholds and Traffic-to-Cashflow Conversion Bounds [短视频完播率风险函数、推荐反馈环与自动化获客漏斗] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090213 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).