关于「钩子插入的动态规划:K 个钩子的最优秒数」,《直播间留存的生存分析》给出的定量结论是什么?

结论:第六章用动态规划排三个钩子,收益 1970.2,均匀每 20 分钟一个平均只有 257.9;节目单推迟一分钟而钩子不动,收益跌到 210.7;没有尖峰时排程只好 1.29 倍,尖峰极高时好 9.53 倍。

数据口径:书中的直播间、观众数、风险参数、钩子效果与节目单均为示意;全部计算由书中程序生成;本书是非个人化经营建议,钩子只能是真实的内容或写明真实库存与真实截止时刻的福利,不用会重置的倒计时,不诱导未成年人消费,平台规则以官方公示为准。 第一章说明截尾让平均停留偏低:示意直播间里把下播当离开得 408.1 秒,只算已离开者得 278.0 秒,乘积极限估计的限定平均停留是 337.9 秒。

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: 1970.2, 20, 257.9

no hook15493.7 audience-minutes per 1000 settled viewers
K1 DP starts 180 gain 1061.4 (re-evaluated 1061.4)
K2 DP starts 180,900 gain 1588.5 (re-evaluated 1588.5)
K3 DP starts 180,900,1260 gain 1970.2 (re-evaluated 1970.2)
K1 brute force over all 360 starts: 180 s, gain 1061.4
K3 uniform every 20 min, 120 phases: mean 257.9 min 79.0 max 1130.1
K3 on the quarter hours 900,1800,2700: gain 579.5
K3 reacting after 3 switches 930,2010,2730: gain 118.9

核心公式

\[V(t) = a_t + e^{-x_t}\, V(t+1).\]
查看完整验算程序
# Chapter 6. p61: dynamic programming for K hooks.
# Programme clock in 10-second bins (360 bins). Hazard per bin from the
# two M1 levels of p41 (rest 0.00072, switch windows 0.0069 per s).
# A hook lasts W = 6 bins at hazard x (1 - rho), rho = 0.5; then for
# E = 3 bins it returns pi = 0.3 of the hazard it removed; two hook
# starts must be at least G = 18 bins (3 min) apart. Objective: audience
# seconds of a closed group of 1000 settled viewers. Because survival
# multiplies, V(t) = area(t) + exp(-x_t) * V(t + 1) is an exact Bellman
# recursion; the state is (bin, hooks used, bins since last start).
import math

T_END, BIN = 3600, 10
NB = T_END // BIN
SWITCH = [180, 900, 1260, 1980, 2700]   # 环节切换秒数(示意)
SPIKE_LEN = 30
REST, SPIKE = 0.00072, 0.0069
W, E, G, RHO, PI = 6, 3, 18, 0.5, 0.3   # 钩子参数(示意)
CAP = max(G, W + E)


def bin_hazard():
    sec = [REST] * T_END
    for s in SWITCH:
        for t in range(s, s + SPIKE_LEN):
            sec[t] = SPIKE
    return [math.fsum(sec[b * BIN:(b + 1) * BIN]) for b in range(NB)]


HB = bin_hazard()
HWIN = [math.fsum(HB[a:min(a + W, NB)]) for a in range(NB)]


def haz(t, c):
    # c = 距上一次钩子开始的箱数(CAP 表示很久以前或从未)
    if c < W:
        return HB[t] * (1.0 - RHO)
    if c < W + E:
        return HB[t] + PI * RHO * HWIN[t - c] / E
    return HB[t]


def step(x):
    return BIN * (1.0 - math.exp(-x)) / x, math.exp(-x)


def dp(k_max):
    nxt = {(k, c): 0.0 for k in range(k_max + 1) for c in range(CAP + 1)}
    pol = []
    for t in range(NB - 1, -1, -1):
        cur, choice = {}, {}
        for k in range(k_max + 1):
            for c in range(CAP + 1):
                opts = [(c, k, False)]
                if c >= G and k < k_max:
                    opts.append((0, k + 1, True))
                best = None
                for ct, kk, start in opts:
                    a, g = step(haz(t, ct))
                    v = a + g * nxt[(kk, min(ct + 1, CAP))]
                    if best is None or v > best[0]:
                        best = (v, start)
                cur[(k, c)], choice[(k, c)] = best
        pol.append(choice)
        nxt = cur
    pol.reverse()
    k, c, starts = 0, CAP, []
    for t in range(NB):
        if pol[t][(k, c)]:
            starts.append(t * BIN)
            k, c = k + 1, 0
        c = min(c + 1, CAP)
    return nxt[(0, CAP)], starts


def forward(starts):
    sb = [s // BIN for s in starts]
    parts, surv, c = [], 1.0, CAP
    for t in range(NB):
        c = 0 if t in sb else min(c + 1, CAP) if t else CAP
        a, g = step(haz(t, c))
        parts.append(surv * a)
        surv *= g
    return math.fsum(parts)


def gain(starts):
    return 1000.0 * (forward(starts) - forward([])) / 60.0


base = 1000.0 * forward([]) / 60.0
print("no hook: %.1f audience-minutes per 1000 settled viewers" % base)
for k in (1, 2, 3):
    v, starts = dp(k)
    print("K=%d DP starts %-18s gain %6.1f  (re-evaluated %6.1f)"
          % (k, ",".join("%d" % s for s in starts),
             1000.0 * (v - forward([])) / 60.0, gain(starts)))
brute = max((gain([a * BIN]), a * BIN) for a in range(NB))
print("K=1 brute force over all 360 starts: %d s, gain %.1f"
      % (brute[1], brute[0]))
uni = [gain([ph, ph + 1200, ph + 2400]) for ph in range(0, 1200, BIN)]
print("K=3 uniform every 20 min, 120 phases: mean %.1f  min %.1f  max %.1f"
      % (math.fsum(uni) / len(uni), min(uni), max(uni)))
print("K=3 on the quarter hours 900,1800,2700: gain %.1f"
      % gain([900, 1800, 2700]))
print("K=3 reacting after 3 switches 930,2010,2730: gain %.1f"
      % gain([930, 2010, 2730]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《直播间留存卡普兰迈耶生存曲线、划走风险尖峰定位与福利钩子插入秒数母典》第 6 章「钩子插入的动态规划:K 个钩子的最优秒数」(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-277 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026277,
  author    = {Fan, Yafu},
  title     = {Kaplan-Meier Retention Curves for Livestream Rooms: Locating Swipe-Away Hazard Spikes and Timing the Insertion of Hooks},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论