关于「像素冷启动,汤普森抽样」,《暗投授权与钩子归因》给出的定量结论是什么?

结论:第四章用汤普森抽样做冷启动,三千点击后挑对最优素材比例只有 0.765。

数据口径:书中参数均为示意,全部计算由书中程序生成;本书是非个人化经营分析,不构成经营或法律建议。 第六章用二叉树为跳单期权定价,合理预估 20.82 已拦住示意模型中全部跳单。

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: 0.765

oracle orders 90; 200 runs of 3,000 clicks
policymean ordersregretclicks on bestfinal pick best
split73.3616.640.3330.710
commit77.2112.790.3930.400
thompson78.0012.000.5200.765
查看完整验算程序
# Chapter 4. p41: cold start of three licensed hooks, three policies.
# Three creatives convert a click with probability 0.020, 0.024 and
# 0.030 (illustrative). Each run spends 3,000 clicks.
#   split     round robin, 1,000 clicks each
#   commit    300 test clicks (100 each), then all on the leader
#   thompson  Beta(1, 1) priors; each click goes to the arm whose
#             posterior draw is largest (Thompson 1933)
# 200 runs with a fixed seed; regret is measured against always
# using the best creative (expected 90 orders).
import math
import random

P = (0.020, 0.024, 0.030)
CLICKS, RUNS = 3000, 200
rng = random.Random(1933)


def split():
    s = [0, 0, 0]
    won = 0
    for c in range(CLICKS):
        a = c % 3
        if rng.random() < P[a]:
            s[a] += 1
            won += 1
    lead = max(range(3), key=lambda i: (s[i], -i))
    return won, CLICKS // 3, lead


def commit():
    s = [0, 0, 0]
    won = 0
    for c in range(300):
        a = c % 3
        if rng.random() < P[a]:
            s[a] += 1
            won += 1
    lead = max(range(3), key=lambda i: (s[i], -i))
    for _ in range(CLICKS - 300):
        if rng.random() < P[lead]:
            won += 1
    on_best = 100 + (CLICKS - 300 if lead == 2 else 0)
    return won, on_best, lead


def thompson():
    s, f = [1, 1, 1], [1, 1, 1]
    won, on_best = 0, 0
    for _ in range(CLICKS):
        draws = [rng.betavariate(s[i], f[i]) for i in range(3)]
        a = max(range(3), key=lambda i: draws[i])
        if a == 2:
            on_best += 1
        if rng.random() < P[a]:
            s[a] += 1
            won += 1
        else:
            f[a] += 1
    lead = max(range(3), key=lambda i: s[i] / (s[i] + f[i]))
    return won, on_best, lead


if __name__ == "__main__":
    oracle = CLICKS * P[2]
    print("oracle orders {:.0f}; {} runs of {:,} clicks".format(
        oracle, RUNS, CLICKS))
    print("  policy    mean orders   regret   clicks on best"
          "   final pick best")
    for name, pol in (("split", split), ("commit", commit),
                      ("thompson", thompson)):
        wins, best, right = [], [], 0
        for _ in range(RUNS):
            w, b, lead = pol()
            wins.append(w)
            best.append(b)
            if lead == 2:
                right += 1
        mw = math.fsum(wins) / RUNS
        mb = math.fsum(best) / RUNS / CLICKS
        print("  {:<8}  {:>11.2f}   {:>6.2f}   {:>14.3f}   {:>15.3f}"
              .format(name, mw, oracle - mw, mb, right / RUNS))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《暗投白名单授权经济学、三秒钩子生存分析与单层星探归因博弈母典》第 4 章「像素冷启动,汤普森抽样」(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-333 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026333,
  author    = {Fan, Yafu},
  title     = {Dark-Ads Whitelist Licensing Economics, Three-Second Hook Survival Analysis and Single-Tier Scout Attribution Games},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论