5D Library › 分销网络幂律激活、网红矩阵组合优化与激励稳定边界 · English

第四章 在投放中学习:汤普森抽样与 CAC 的尾部

先试一周,还是边投边学

第三章假定每位达人的获客能力已知。实际上要靠投放去学:四十位中小达人,每位每次点击的真实转化率不同(中位数 1%,差异很大),每周有固定数量的付费点击要分给他们,共十二周。三种规则:平均分配;第一周平均、之后全部给目前表现最好的人(贪心);以及汤普森抽样,按每位达人「是最好的人」的后验概率分配点击 (Thompson 1933; Russo et al. 2018)。

推导

# Chapter 4. p07: Learning which micro creators convert. 40 creators;
# each has a true conversion rate per click, lognormal with median 1%
# and log sd 0.8 (capped at 10%). Each week 2,000 or 8,000 paid clicks
# are split among them, for 12 weeks. Uniform: equal shares every
# week. Greedy: uniform in week 1, then everything to the best rate seen
# so far. Thompson: each creator's rate has a Beta(1 + sales, 1 + misses)
# posterior; clicks go in proportion to how often (in 100 draws) each
# creator's draw is the highest. Oracle: all clicks to the true best.
# Mean and 10th percentile over 150 repeats, seed 87.
import random
from math import exp, sqrt

K, WEEKS = 40, 12

def count(rng, n, p):
    lam = n * p
    if lam > 30:
        return max(0, round(rng.gauss(lam, sqrt(lam * (1 - p)))))
    k, t, u = 0, exp(-lam), rng.random()
    c = t
    while u > c:
        k += 1
        t *= lam / k
        c += t
    return k

def world(rng):
    return [min(0.10, 0.01 * exp(0.8 * rng.gauss(0, 1))) for _ in range(K)]

def play(rng, rates, rule, clicks):
    s, f, total = [0] * K, [0] * K, 0
    for week in range(WEEKS):
        if rule == "uniform" or (rule == "greedy" and week == 0):
            share = [1.0 / K] * K
        elif rule == "greedy":
            best = max(range(K), key=lambda i: (s[i] + 1) / (s[i] + f[i] + 2))
            share = [1.0 if i == best else 0.0 for i in range(K)]
        else:
            wins = [0] * K
            for _ in range(100):
                d = [rng.betavariate(1 + s[i], 1 + f[i]) for i in range(K)]
                wins[d.index(max(d))] += 1
            share = [w / 100 for w in wins]
        for i in range(K):
            n = int(round(clicks * share[i]))
            if n:
                c = count(rng, n, rates[i])
                s[i] += c
                f[i] += n - c
                total += c
    return total

if __name__ == "__main__":
    rng = random.Random(87)
    print("   clicks a week  rule      share of oracle  10th percentile")
    for clicks in (2000, 8000):
        res = {r: [] for r in ("uniform", "greedy", "thompson")}
        for _ in range(150):
            rates = world(rng)
            o = max(rates) * clicks * WEEKS
            for r in res:
                res[r].append(play(rng, rates, r, clicks) / o)
        for r, v in res.items():
            v.sort()
            print("   {:14,d}  {:<8}  {:15.3f}  {:15.3f}".format(
                clicks, r, sum(v) / 150, v[15]))
#    clicks a week  rule      share of oracle  10th percentile
#             2,000  uniform             0.257            0.165
#             2,000  greedy              0.674            0.425
#             2,000  thompson            0.663            0.573
#             8,000  uniform             0.251            0.173
#             8,000  greedy              0.891            0.803
#             8,000  thompson            0.822            0.776

平均分配只拿到理想情况(全部给真正最好的人)的四分之一左右(0.257 与 0.251)。每周 8,000 次点击时,贪心规则平均拿到 0.891,汤普森抽样 0.822,贪心反而略好:第一周每人分到两百次点击,已足以大致认出最好的人。每周只有 2,000 次点击时,两者的平均相近(0.674 与 0.663),但最差的一成情况差别明显:贪心只有 0.425,汤普森抽样有 0.573。预算小、第一周的信息不够时,贪心容易押错人,汤普森抽样的价值在于保护下限。

报告哪一个 CAC

每场中小达人活动投入 2,500 令吉,获得的客户数是一个泊松计数,其均值本身差异很大。p08 比较一年 384 场活动的几种 CAC 报告方式。

推导

# Chapter 4. p08: Which CAC to report. Each micro campaign spends
# RM 2,500 and wins a Poisson number of customers whose mean is
# lognormal (median 5, log sd 1.0). Over a year of 32 campaigns a month
# (384 campaigns), compare the share of campaigns with no customer at
# all, the mean and the median of per-campaign CAC (over campaigns with
# at least one customer), and the pooled CAC = total spend / total
# customers, and the share of campaigns that bring in half of all
# customers. 1,000 simulated years, seed 87.
import random
from math import exp

def poisson(rng, lam):
    k, t, u = 0, exp(-lam), rng.random()
    c = t
    while u > c and k < 1000:
        k += 1
        t *= lam / k
        c += t
    return k

def year(rng):
    wins = [poisson(rng, 5.0 * exp(rng.gauss(0, 1.0))) for _ in range(384)]
    cac = sorted(2500.0 / w for w in wins if w)
    zero = 1 - len(cac) / 384
    pooled = 2500.0 * 384 / sum(wins)
    acc, half = 0, 0
    for w in sorted(wins, reverse=True):
        acc += w
        half += 1
        if acc >= sum(wins) / 2:
            break
    return zero, sum(cac) / len(cac), cac[len(cac) // 2], pooled, half / 384

if __name__ == "__main__":
    rng = random.Random(87)
    ys = [year(rng) for _ in range(1000)]
    names = ("share with no customer", "mean per-campaign CAC (RM)",
             "median per-campaign CAC (RM)", "pooled CAC (RM)",
             "campaigns giving half of sales")
    print("   measure                          median   5% to 95% of years")
    for j, name in enumerate(names):
        v = sorted(y[j] for y in ys)
        fmt = "{:.3f}" if j in (0, 4) else "{:,.0f}"
        print("   {:<31}  {:>7}   {} to {}".format(
            name, fmt.format(v[500]), fmt.format(v[50]), fmt.format(v[949])))
#    measure                          median   5% to 95% of years
#    share with no customer             0.073   0.049 to 0.094
#    mean per-campaign CAC (RM)           723   663 to 789
#    median per-campaign CAC (RM)         500   417 to 500
#    pooled CAC (RM)                      304   271 to 341
#    campaigns giving half of sales     0.154   0.130 to 0.174

约 0.073 的活动一个客户都没有,它们的 CAC 是无穷大,只能被排除在「平均每场 CAC」之外。即便如此,平均每场 CAC 是 723 令吉,每场 CAC 的中位数是 500 令吉,而真正决定盈亏的合并 CAC(总投入除以总客户)只有 304 令吉。三个数相差一倍以上,却都可以被叫作「CAC」。另外,0.154 的活动带来了一半的客户:活动层面的产出同样是重尾的。

注

CAC 的报告规矩: 以合并 CAC(总投入除以总客户)作为主要指标;另报零客户活动的比例与带来一半客户的活动比例;不要对每场活动的 CAC 取平均。

小结

在投放中学习时,预算充足则「先平均试一周再集中」已经不错;预算小时,汤普森抽样能明显改善最差情况。CAC 有三种常见算法,彼此相差一倍以上;只有合并 CAC 对应真实的盈亏。

参考文献

Russo, Daniel J., Benjamin Van Roy, Abbas Kazerouni, Ian Osband, and Zheng Wen. 2018. “A Tutorial on Thompson Sampling.” Foundations and Trends in Machine Learning 11 (1): 1 to 96.
Thompson, William R. 1933. “On the Likelihood That One Unknown Probability Exceeds Another in View of the Evidence of Two Samples.” Biometrika 25 (3/4): 285 to 294.

Fan, Y. (2026). Affiliate Network Power Laws, Influencer Portfolio Optimization and Incentive Stability Bounds: Heavy-Tailed Productivity, Markowitz Traffic Allocation and Single-Tier Commission Equilibria [分销网络幂律激活、网红矩阵组合优化与激励稳定边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090229 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).