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

第一章 划走的动力学:威布尔风险函数

把划走写成风险函数

一条短视频开始播放后,观众随时可能划走。设观众停留的时间为随机变量 TT,留存函数 S(t)S(t) 是看到第 tt 秒仍未离开的比例,风险函数 h(t)h(t) 是已经看到第 tt 秒的人在下一瞬间离开的速率。生存分析的标准做法是给 TT 选一个分布 (Klein and Moeschberger 2003);威布尔分布 (Weibull 1951) 的留存与风险为

S(t)=e−(t/λ)κ,h(t)=κλ(tλ)κ−1.S(t) = e^{-(t/\lambda)^{\kappa}},\qquad h(t) = \frac{\kappa}{\lambda}\Big(\frac{t}{\lambda}\Big)^{\kappa-1}.

形状参数 κ<1\kappa < 1 时风险随时间递减:划走集中在开头,留下来的人越来越不容易离开。这正是短视频的典型形态。κ=1\kappa = 1 时退化为风险恒定的指数分布。

推导

# Chapter 1. p01: Weibull swipe-away kinetics and completion rates.
# The time a viewer stays on a short video is modelled as Weibull with
# shape k and scale L (seconds): survival S(t) = exp(-(t / L)^k),
# hazard h(t) = (k / L) (t / L)^(k - 1). With k < 1 the hazard falls
# over time: most swipes come in the first seconds, and those who stay
# become ever less likely to leave. For a video of length D:
#   completion rate = S(D);
#   mean watch time = integral of S(t) from 0 to D (midpoint rule);
#   watch fraction = mean watch time / D.
# An exponential model (k = 1) tuned to the same 5-second retention is
# shown for contrast. Last, k and L are recovered from two retention
# readings, since ln(-ln S) = k ln t - k ln L is a straight line.
import math

def surv(t, k, L):
    return math.exp(-(t / L) ** k)

def haz(t, k, L):
    return (k / L) * (t / L) ** (k - 1)

def mean_watch(D, k, L, n=4000):
    return sum(surv((i + 0.5) * D / n, k, L) for i in range(n)) * D / n

if __name__ == "__main__":
    k, L = 0.6, 20.0
    s5 = surv(5, k, L)
    Le = -5 / math.log(s5)
    print("   Weibull k = {}, L = {} s".format(k, L))
    print("   retention at 3 s {:.3f}, at 5 s {:.3f}".format(
        surv(3, k, L), s5))
    print("   hazard per second at 1, 5, 20 s: {:.3f}, {:.3f}, {:.3f}".format(
        haz(1, k, L), haz(5, k, L), haz(20, k, L)))
    print("   exponential with the same 5 s retention: scale {:.2f} s"
          .format(Le))
    print("\n   length D   completion   mean watch   watch frac   exp compl")
    for D in (7, 15, 30, 60, 90):
        mw = mean_watch(D, k, L)
        print("   {:6d} s {:12.3f} {:10.1f} s {:12.3f} {:11.3f}".format(
            D, surv(D, k, L), mw, mw / D, surv(D, 1.0, Le)))
    t1, t2 = 3.0, 15.0
    a1 = math.log(-math.log(surv(t1, k, L)))
    a2 = math.log(-math.log(surv(t2, k, L)))
    kk = (a2 - a1) / (math.log(t2) - math.log(t1))
    LL = math.exp(math.log(t1) - a1 / kk)
    print("\n   recovered from retention at 3 s and 15 s:")
    print("   k = {:.3f}, L = {:.2f}".format(kk, LL))
    print("\nReading: with a falling hazard, a 60 s video keeps far more")
    print("viewers to the end than an exponential model with the same")
    print("first five seconds predicts. Completion rate punishes length,")
    print("but much less than intuition suggests; two retention points")
    print("are enough to fit both parameters.")
#    Weibull k = 0.6, L = 20.0 s
#    retention at 3 s 0.726, at 5 s 0.647
#    hazard per second at 1, 5, 20 s: 0.099, 0.052, 0.030
#    exponential with the same 5 s retention: scale 11.49 s
#
#    length D   completion   mean watch   watch frac   exp compl
#         7 s        0.587        5.1 s        0.723       0.544
#        15 s        0.431        9.1 s        0.605       0.271
#        30 s        0.279       14.3 s        0.476       0.073
#        60 s        0.145       20.3 s        0.339       0.005
#        90 s        0.085       23.7 s        0.263       0.000
#
#    recovered from retention at 3 s and 15 s:
#    k = 0.600, L = 20.00
#
# Reading: with a falling hazard, a 60 s video keeps far more
# viewers to the end than an exponential model with the same
# first five seconds predicts. Completion rate punishes length,
# but much less than intuition suggests; two retention points
# are enough to fit both parameters.

取 κ=0.6\kappa = 0.6、λ=20\lambda = 20 秒,3 秒留存为 0.726,5 秒留存为 0.647;第 1 秒的风险是每秒 0.099,第 20 秒降到 0.030。15 秒视频的完播率为 0.431,60 秒为 0.145;若用 5 秒留存相同的指数模型(尺度 11.49 秒),60 秒的完播率只有 0.005。递减风险意味着:完播率确实惩罚时长,但远没有直觉那么严厉。平均观看时长随时长增长而增长,60 秒视频平均被看 20.3 秒,观看比例却降到 0.339。最后,程序只用 3 秒与 15 秒两个留存读数,就恢复了 κ=0.600\kappa = 0.600 与 λ=20.00\lambda = 20.00:因为 ln⁡(−ln⁡S)\ln(-\ln S) 对 ln⁡t\ln t 是一条直线。

威布尔留存(实线)与 5 秒留存相同的指数留存(虚线)。

中途再钩值不值得

「每隔几秒就要再钩一次」是常见的剪辑建议。程序假设每次再钩使随后 2 秒的风险降低七成,但观众会学会、会厌倦:每分钟 nn 次再钩使整条视频的风险提高 cn2c\,n^2 的比例。

推导

# Chapter 1. p02: Re-hook pulses: how often should a video re-hook?
# Baseline hazard: Weibull k = 0.6, L = 20 s, as in p01, for a 60 s
# video. A re-hook (a question, a cut, a promise of what comes next)
# placed every H seconds lowers the hazard by a factor (1 - r) for the
# w = 2 s that follow it. Re-hooks also wear: with n hooks per minute
# the whole hazard rises by a share c n^2 (viewers learn the trick and
# tire of it). With r = 0.7 and c = 0.002 the program integrates the
# cumulative hazard on a 0.01 s grid and reports completion and mean
# watch time.
import math

K, L, D = 0.6, 20.0, 60.0

def base(t):
    return (K / L) * (t / L) ** (K - 1)

def run(H, r=0.7, w=2.0, c=0.002, dt=0.01):
    per_min = 0.0 if H is None else 60.0 / H
    lift = 1 + c * per_min ** 2
    cum, watch, t = 0.0, 0.0, 0.0
    while t < D:
        tm = t + dt / 2
        h = base(tm) * lift
        if H is not None and tm >= H and (tm % H) < w:
            h *= 1 - r
        watch += math.exp(-cum - h * dt / 2) * dt
        cum += h * dt
        t += dt
    return math.exp(-cum), watch

if __name__ == "__main__":
    print("   hook every   hooks/min   completion   mean watch (s)")
    for H in (None, 30, 20, 15, 10, 6, 4, 3):
        comp, mw = run(H)
        print("   {:>10s} {:11.1f} {:12.3f} {:14.1f}".format(
            "none" if H is None else "{} s".format(H),
            0.0 if H is None else 60.0 / H, comp, mw))
    print("\nReading: re-hooks after the opening buy only a little,")
    print("because with a falling hazard most viewers who will swipe")
    print("have already swiped before the first re-hook arrives. Hooks")
    print("closer than every 10 s lose ground, and at every 4 s the")
    print("video does worse than with none.")
    print("The opening seconds, not the middle, decide completion.")
#    hook every   hooks/min   completion   mean watch (s)
#          none         0.0        0.145           20.4
#          30 s         2.0        0.148           20.4
#          20 s         3.0        0.151           20.5
#          15 s         4.0        0.153           20.5
#          10 s         6.0        0.155           20.5
#           6 s        10.0        0.150           19.9
#           4 s        15.0        0.136           18.4
#           3 s        20.0        0.122           16.8
#
# Reading: re-hooks after the opening buy only a little,
# because with a falling hazard most viewers who will swipe
# have already swiped before the first re-hook arrives. Hooks
# closer than every 10 s lose ground, and at every 4 s the
# video does worse than with none.
# The opening seconds, not the middle, decide completion.

没有再钩时完播率为 0.145;每 10 秒一次时达到最高的 0.155;更密就开始倒退,每 4 秒一次时为 0.136,已不如不钩。增益很小的原因与第一节相同:在递减风险下,会划走的人大多在第一次再钩出现之前就已经划走了。开头几秒,而不是中段,决定了完播率。

假说 · 尚未验证

可检验的推论。 若划走风险确实随时间递减,则同一账号不同视频的留存曲线在 ln⁡(−ln⁡S)\ln(-\ln S) 对 ln⁡t\ln t 的坐标上应近似为直线,斜率小于 1;改进开头的剪辑应显著提高 5 秒留存,而在 20 秒之后加入再钩只能带来很小的完播率变化。什么会证明这一推论错:若留存曲线的斜率接近或大于 1,或中段再钩带来与开头改进相当的提升,本章的模型就不适用于该账号。

小结

划走可以写成威布尔风险函数;短视频的风险随时间递减,所以开头决定一切,而完播率对时长的惩罚比指数模型温和得多。两个留存读数就能拟合两个参数。中途再钩的收益有限,太密则有害。

参考文献

Klein, John P., and Melvin L. Moeschberger. 2003. Survival Analysis: Techniques for Censored and Truncated Data. 2nd ed. Springer.
Weibull, Waloddi. 1951. “A Statistical Distribution Function of Wide Applicability.” Journal of Applied Mechanics 18 (3): 293–97.

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).