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
第二章 分级流量池与第一关的误杀
一个程式化的流量池阶梯
关于推荐系统,创作者之间流传着一个简化的图景:每条新视频先推给约 300 人;表现排在前列的晋级到 3,000 人的池子,再晋级到 50,000 人、1,000,000 人。真实平台的推荐系统要复杂得多,通常是从海量候选中先召回、再排序的多级神经网络 (Covington et al. 2016)。本章不声称描述任何平台,只问一个算术问题:如果晋级依据的是小样本上观测到的完播率,会发生什么?
程序假设各视频的真实完播率服从均值 0.25 的 Beta 分布,每一级只让观测完播率前 10% 的视频晋级;每一级的观众与上一级不同,在 logit 尺度上有一个标准差为 的「观众差异」。
推导
# Chapter 2. p03: How often does a good video die in the first pool?
# A stylised traffic-pool ladder: every new video is shown to 300
# viewers; the top 10 % by observed completion go on to 3,000, the top
# 10 % of those to 50,000, the top 10 % of those to 1,000,000. True
# completion rates p across videos follow Beta(5, 15) (mean 0.25).
# Each pool's audience also differs from the last: on the logit scale
# the rate a pool sees is shifted by N(0, s), the "audience draw".
# Observed rate = binomial(N, p') / N, normal approximation, clipped.
# "Good" means a true rate in the top 10 % of all videos; "best" means
# the top 0.1 %, as many videos as there are places in the last pool.
# 200,000 videos, fixed seed. False negative = a good video stopped.
import math
import random
STAGES = [300, 3000, 50000]
def logit(p):
return math.log(p / (1 - p))
def expit(x):
return 1 / (1 + math.exp(-x))
def run(s, n=200000, seed=8):
rng = random.Random(seed)
vids = [rng.betavariate(5, 15) for _ in range(n)]
alive = list(range(n))
reached = [alive]
for N in STAGES:
obs = []
for i in alive:
p = expit(logit(vids[i]) + rng.gauss(0, s)) if s else vids[i]
x = p + rng.gauss(0, 1) * math.sqrt(p * (1 - p) / N)
obs.append((x, i))
obs.sort(reverse=True)
alive = [i for _, i in obs[:len(obs) // 10]]
reached.append(alive)
srt = sorted(vids, reverse=True)
t10, t1 = srt[n // 10], srt[n // 1000]
good10 = [i for i in range(n) if vids[i] > t10]
good1 = [i for i in range(n) if vids[i] > t1]
s1 = set(reached[1])
s3 = set(reached[3])
fn1 = sum(1 for i in good10 if i not in s1) / len(good10)
top1_to_end = sum(1 for i in good1 if i in s3) / len(good1)
prec_end = sum(1 for i in s3 if vids[i] > t1) / len(s3)
return t10, fn1, top1_to_end, prec_end, len(s3)
if __name__ == "__main__":
print(" audience draw s good videos stopped best videos reaching")
print(" at 300 1,000,000")
for s in (0.0, 0.2, 0.4):
t10, fn1, reach, prec, m = run(s)
print(" {:15.1f} {:21.3f} {:22.3f}".format(s, fn1, reach))
print(" (top 10 % threshold on true completion {:.3f}; {} videos"
.format(t10, m))
print(" reach the last pool, so the share of best videos reaching")
print(" it is also the share of the last pool that is best)")
print("\nReading: with 300 viewers the observed rate has a standard")
print("error of about 0.025, comparable to the gap between a good")
print("and an average video, so 0.169 of good videos stop at the")
print("first gate even when every audience is alike. Audience")
print("variation between pools raises the loss to 0.308 or 0.478,")
print("and then only about a half or a quarter of the best videos")
print("reach the top pool.")
# audience draw s good videos stopped best videos reaching
# at 300 1,000,000
# 0.0 0.169 0.970
# 0.2 0.308 0.525
# 0.4 0.478 0.245
# (top 10 % threshold on true completion 0.378; 200 videos
# reach the last pool, so the share of best videos reaching
# it is also the share of the last pool that is best)
#
# Reading: with 300 viewers the observed rate has a standard
# error of about 0.025, comparable to the gap between a good
# and an average video, so 0.169 of good videos stop at the
# first gate even when every audience is alike. Audience
# variation between pools raises the loss to 0.308 or 0.478,
# and then only about a half or a quarter of the best videos
# reach the top pool.300 人样本下观测完播率的标准误约为 0.025,与好视频和普通视频之间的差距相当。即使每一级的观众完全相同,真实前 10% 的好视频也有 0.169 在第一关被拦下;观众差异为 0.2 时为 0.308,为 0.4 时为 0.478。最终进入百万级池子的只有 200 条视频;观众相同时其中 0.970 是真实的前 0.1%,观众差异为 0.2 和 0.4 时只剩 0.525 和 0.245。也就是说,进入顶层的大多是好视频,但大多数好视频到不了顶层。流传的「第一关误杀三到四成」在观众差异为 0.2 至 0.4 之间时与本模型相符。
重发与重剪:对冲第一关的噪声
既然第一关是一道有噪声的门,创作者自然会想到把同一条视频再发一次。设每次尝试独立地以概率 过关, 次尝试至少过一次的概率为 ;但重发的视频以概率 被识别为重复内容,一旦被识别,账号要付出相当于 1.5 次过关价值的代价。重剪(换开头、换封面)被当作新视频,不会被识别,但每次要花 0.1 的制作成本。
推导
# Chapter 2. p04: Reposting the same video: a hedging matrix.
# If one test pool is a noisy gate, a creator may post the same video
# again. Each attempt passes the first gate with chance q, independent
# of the others, so k attempts pass at least once with 1 - (1 - q)^k.
# A pass is worth V = 1. But each repost after the first is caught as
# a duplicate with chance d, and a catch costs the account P = 1.5
# (a traffic cut that lasts for weeks). Expected value of k attempts:
# EV = 1 - (1 - q)^k - P * (1 - (1 - d)^(k - 1)).
# A re-edit (new opening, new cover) is treated as a fresh video: it
# is never flagged but costs E = 0.1 per attempt to make.
# q = 0.6 for a good video, 0.05 for an average one (illustrative).
P, E = 1.5, 0.1
def ev_repost(q, k, d):
return 1 - (1 - q) ** k - P * (1 - (1 - d) ** (k - 1))
def ev_reedit(q, k):
return 1 - (1 - q) ** k - E * (k - 1)
if __name__ == "__main__":
for label, q in (("good video, q = 0.6", 0.6),
("average video, q = 0.05", 0.05)):
print(" " + label)
print(" attempts k d = 0 d = 0.2 d = 0.5 d = 0.8"
" re-edit")
for k in (1, 2, 3, 4):
row = [ev_repost(q, k, d) for d in (0.0, 0.2, 0.5, 0.8)]
print(" {:10d}".format(k) + "".join(
"{:9.3f} ".format(v) for v in row).rstrip() +
"{:10.3f}".format(ev_reedit(q, k)))
print()
print("Reading: when duplicates are never caught, reposting a good")
print("video is a cheap hedge against the noisy first gate. Once the")
print("platform catches a fifth of reposts, a second post already")
print("loses value, and one re-edit (0.740) is the better hedge.")
print("For an average video re-edits never pay and reposts pay only")
print("if nothing is ever caught. The fix is a better video.")
# good video, q = 0.6
# attempts k d = 0 d = 0.2 d = 0.5 d = 0.8 re-edit
# 1 0.600 0.600 0.600 0.600 0.600
# 2 0.840 0.540 0.090 -0.360 0.740
# 3 0.936 0.396 -0.189 -0.504 0.736
# 4 0.974 0.242 -0.338 -0.514 0.674
#
# average video, q = 0.05
# attempts k d = 0 d = 0.2 d = 0.5 d = 0.8 re-edit
# 1 0.050 0.050 0.050 0.050 0.050
# 2 0.098 -0.202 -0.652 -1.103 -0.002
# 3 0.143 -0.397 -0.982 -1.297 -0.057
# 4 0.185 -0.547 -1.127 -1.303 -0.115
#
# Reading: when duplicates are never caught, reposting a good
# video is a cheap hedge against the noisy first gate. Once the
# platform catches a fifth of reposts, a second post already
# loses value, and one re-edit (0.740) is the better hedge.
# For an average video re-edits never pay and reposts pay only
# if nothing is ever caught. The fix is a better video.对好视频(),重复内容从不被识别时,重发是廉价的对冲:两次尝试的期望值从 0.600 升到 0.840。但只要平台识别出五分之一的重发,第二次发布就已经亏损(0.540);这时重剪一次(0.740)是更好的对冲。对普通视频(),重剪永远不划算,重发只有在从不被识别时才划算。真正的办法是做一条更好的视频。
注
操作上的含义。 第一关没过,先看 3 秒与 5 秒留存是否明显低于账号的常态:若是,问题在开头,重剪开头;若不是,视频很可能只是被噪声拦下,可以重剪后再试一次。不要原样重发。
小结
在小样本的晋级门槛下,好视频被误杀是常态而不是例外;观众差异越大,误杀越多。对冲噪声的正确方式是重剪而不是重发,而对普通视频,任何对冲都不如改进内容。
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).