关于「按浏览行为判别加工风格:贝叶斯更新盖过性别先验」,《加工广度与后悔安全垫》给出的定量结论是什么?

结论:第三章只用站内四类行为计数做贝叶斯判别,模拟准确率 0.864,加回性别先验几乎不再提升。

数据口径:梅耶斯-列维的选择性加工假说讲的不是感性对理性,而是加工广度:启发式加工者看少数显著线索,综合加工者把全部线索都算进去,线索彼此一致或过弱过强时差异消失。 海德的性别相似性假说与后续元合成显示,多数心理变量的性别差异很小,组内差异远大于组间差异;d 为 0.30 时,按性别猜加工风格的最好命中率只有 0.560,按性别分流要追平一张分层页面,d 须达到 1.79。

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: 0.864

rule accuracy Brier (20,000 simulated sessions)
G0.561 0.246
B0.864 0.096
BG0.868 0.095
behaviour-only accuracy by number of signals observed
1 signal(s)0.784
2 signal(s)0.824
3 signal(s)0.858
4 signal(s)0.864
gain from adding the gender prior: +0.0035
pre-registered floor for behaviour-only accuracy0.65
查看完整验算程序
# Chapter 3. p32: behaviour beats a gender prior, by how much, and when.
# Simulated sessions (fixed seed): a visitor's style is comprehensive with
# probability 0.56 in group A and 0.44 in group B (d = 0.3 read as shares,
# illustrative). Signals follow the rates of p31. We score three rules:
#   G  gender only: A -> comprehensive, B -> heuristic
#   B  behaviour only, prior 0.5, threshold 0.5
#   BG behaviour with the group prior
# and report accuracy and Brier score, then accuracy by signals observed.
import math
import random

rng = random.Random(37332)
RATES = {1: (4.0, 1.2, 2.0, 6.0), 0: (1.5, 0.4, 0.8, 3.0)}
N = 20000


def pois(lam):
    k, p, L = 0, 1.0, math.exp(-lam)
    while True:
        p *= rng.random()
        if p < L:
            return k
        k += 1


def log_lik(s, style, use):
    r = RATES[style]
    parts = []
    for j in range(use):
        if j < 3:
            parts.append(s[j] * math.log(r[j]) - r[j])
        else:
            parts.append(-math.log(r[3]) - s[3] / r[3])
    return math.fsum(parts)


def post(s, prior, use=4):
    a = math.log(prior) + log_lik(s, 1, use)
    b = math.log(1 - prior) + log_lik(s, 0, use)
    return 1.0 / (1.0 + math.exp(b - a))


data = []
for i in range(N):
    grp = i % 2
    style = 1 if rng.random() < (0.56 if grp == 0 else 0.44) else 0
    r = RATES[style]
    s = (pois(r[0]), pois(r[1]), pois(r[2]), rng.expovariate(1 / r[3]))
    data.append((grp, style, s))


def score(probs):
    hit, br = 0, []
    for (g, y, s), p in zip(data, probs):
        hit += (p >= 0.5) == (y == 1)
        br.append((p - y) ** 2)
    return hit / N, math.fsum(br) / N


if __name__ == "__main__":
    pg = [0.56 if g == 0 else 0.44 for g, _, _ in data]
    pb = [post(s, 0.5) for _, _, s in data]
    pbg = [post(s, 0.56 if g == 0 else 0.44) for g, _, s in data]
    print("rule  accuracy  Brier   ({:,} simulated sessions)".format(N))
    for name, pr in (("G", pg), ("B", pb), ("BG", pbg)):
        acc, br = score(pr)
        print("  {:<3} {:.3f}    {:.3f}".format(name, acc, br))
    print("behaviour-only accuracy by number of signals observed")
    for use in (1, 2, 3, 4):
        acc, _ = score([post(s, 0.5, use) for _, _, s in data])
        print("  {} signal(s): {:.3f}".format(use, acc))
    print("gain from adding the gender prior: {:+.4f}".format(
        score(pbg)[0] - score(pb)[0]))
    print("pre-registered floor for behaviour-only accuracy: 0.65")

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《性别消费决策异质性、综合信息加工假说与女性高客单转化工程母典》第 3 章「按浏览行为判别加工风格:贝叶斯更新盖过性别先验」(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-373 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026373,
  author    = {Fan, Yafu},
  title     = {Gendered Consumption Decision Heterogeneity, the Selectivity Hypothesis and High-Conversion Architecture for Female Markets},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论