5D Library › 结账顺手加购弹性、一键升单动力学与跨领域推荐矩阵 · English

第六章 加购定价的在线学习与辛普森悖论

平均分流还是汤普森抽样

第一章说最优加购价格要在自己的流量上测出来。最常见的做法是平均分流:四个候选价格各占四分之一的订单,测试结束再选最好的。另一种做法是汤普森抽样:对每个价格的接受率维持一个贝塔后验,每笔订单按后验抽样选出期望收入最高的价格,逐步把流量集中到好的价格上 (Thompson 1933; Russo et al. 2018)。

推导

# Chapter 6. p11: Choosing a bump price by Thompson sampling.
# Four candidate bump prices with true take rates from p01's model at
# zero friction. Revenue per order = price x take. Over 20,000 orders we
# compare an even split (A/B/C/D test, then nothing learned until the
# end) with Thompson sampling on Beta posteriors of take, choosing the
# price with the highest sampled revenue. Regret = revenue lost against
# always using the best price.
import random
from math import log, erf, sqrt

def take(p, med=0.45, sd=0.8):
    return 0.5 * (1 - erf((log(p / 98) - log(med)) / sd / sqrt(2)))

PRICES = (19.0, 29.0, 44.0, 69.0)

def run(policy, n=20000, seed=78011):
    rng = random.Random(seed)
    a = [1.0] * 4
    b = [1.0] * 4
    rev = 0.0
    for t in range(n):
        if policy == "split":
            k = t % 4
        else:
            k = max(range(4), key=lambda i: PRICES[i] * rng.betavariate(
                a[i], b[i]))
        if rng.random() < take(PRICES[k]):
            rev += PRICES[k]
            a[k] += 1
        else:
            b[k] += 1
    return rev

if __name__ == "__main__":
    for p in PRICES:
        print("   bump RM {:4.0f}: take {:.3f}, per order {:.2f}".format(
            p, take(p), p * take(p)))
    best = max(p * take(p) for p in PRICES) * 20000
    for pol in ("split", "thompson"):
        r = run(pol)
        print("   {:8s} revenue RM {:,.0f}; regret RM {:,.0f}".format(
            pol, r, best - r))
#    bump RM   19: take 0.854, per order 16.22
#    bump RM   29: take 0.700, per order 20.30
#    bump RM   44: take 0.501, per order 22.05
#    bump RM   69: take 0.288, per order 19.86
#    split    revenue RM 392,991; regret RM 48,005
#    thompson revenue RM 439,165; regret RM 1,831

在第一章的模型中,四个候选价格 RM 19、29、44、69 的每单加购收入分别为 16.22、20.30、22.05、19.86,最好的是 RM 44。20,000 笔订单中,平均分流比「始终用最好价格」少收 RM 48,005,汤普森抽样只少收 RM 1,831。代价是:汤普森抽样给出的是「把钱赚到」,而不是一个干净的效应估计;若需要报告每个价格的精确接受率,平均分流更合适。

合并的数字会骗人

推导

# Chapter 6. p12: Simpson's paradox in pooled cross-sell numbers.
# Two recommendation layouts, shown to two traffic sources with very
# different baseline interest (returning buyers vs new readers). Layout
# B wins within each source but loses in the pooled total, because it
# was shown mostly to new readers. Counts are illustrative.
DATA = {"A": {"returning": (900, 81), "new": (100, 2)},
        "B": {"returning": (100, 10), "new": (900, 27)}}

if __name__ == "__main__":
    for lay, d in DATA.items():
        for src, (n, k) in d.items():
            print("   layout {} {:9s}: {:4d} shown, {:3d} clicks, rate {:.3f}"
                  .format(lay, src, n, k, k / n))
        n = sum(v[0] for v in d.values())
        k = sum(v[1] for v in d.values())
        print("   layout {} pooled   : {:4d} shown, {:3d} clicks, rate {:.3f}"
              .format(lay, n, k, k / n))
#    layout A returning:  900 shown,  81 clicks, rate 0.090
#    layout A new      :  100 shown,   2 clicks, rate 0.020
#    layout A pooled   : 1000 shown,  83 clicks, rate 0.083
#    layout B returning:  100 shown,  10 clicks, rate 0.100
#    layout B new      :  900 shown,  27 clicks, rate 0.030
#    layout B pooled   : 1000 shown,  37 clicks, rate 0.037

两种推荐版式分别展示给回头客与新读者。版式 B 在回头客中点击率 0.100(A 为 0.090),在新读者中 0.030(A 为 0.020),两组都胜出;但合并之后 B 只有 0.037,A 为 0.083,因为 B 主要展示给了点击率本来就低的新读者。这就是辛普森悖论 (Simpson 1951):比较推荐效果时必须按流量来源分层,或者随机分配。

注

避免这一陷阱的唯一可靠办法是随机化。 若版式由人工或算法按流量类型分配,合并后的比较就同时混入了版式效应与流量差异。

小结

汤普森抽样把定价测试的损失从 RM 48,005 降到 RM 1,831(示意模型),适合以收入为目标的持续优化。推荐效果的比较必须分层或随机化,否则合并数字可能给出与各层相反的结论。

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): 1–96.
Simpson, Edward H. 1951. “The Interpretation of Interaction in Contingency Tables.” Journal of the Royal Statistical Society: Series B 13 (2): 238–41.
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): 285–94.

Fan, Y. (2026). Order-Bump Friction Elasticity, One-Click Upsell Kinetics and Cross-Domain Recommendation Matrices: Zero-Reentry Impulse Attachment, Conditional Mutual Information and Cart-Value Multipliers [结账顺手加购弹性、一键升单动力学与跨领域推荐矩阵] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090207 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).