5D Library › 超临界病毒级联、信息缺口风险率与多层付费墙优化 · English

第四章 信息缺口、购买风险率与「说得准」的感觉

好奇心是一个缺口

罗温斯坦把好奇心定义为对一个「信息缺口」的反应:当一个人意识到自己知道的与想知道的之间存在差距时,就产生一种想把它填上的冲动 (Loewenstein 1994)。这个理论有两个可以检验的推论。第一,什么都不知道时不会好奇——问题还没有形成;几乎什么都知道时也不会——缺口已经合上。好奇心在中间最强,强度与缺口的大小和显著性有关 (Golman and Loewenstein 2018)。第二,好奇心会消退:缺口不被填上,人的注意力会转走。

证据

倒 U 形的好奇心。 康等人让被试回答常识问题并报告自己的把握程度与好奇程度:好奇心在把握中等时最高,把握很低或很高时都较低;好奇心也预测了被试愿意花代价去获得答案 (Kang et al. 2009)。

第七段程序把这两点写成一个购买风险率:免费钩子揭示了完整结果的比例 rr,好奇强度取 C(r)=4r(1−r)C(r) = 4r(1-r),购买风险率随时间以半衰期约 0.35 天衰减。程序中的 1−exp⁡(−h0Cτe−T/τ)1-\exp\!\big(-h_0 C \tau\, e^{-T/\tau}\big) 是付费墙在延迟 TT 弹出之后、时间延伸到无穷时的累计购买转化率。

推导

# Chapter 4. An information gap as a purchase hazard.
# A free teaser reveals a share r of what the full reading will say. In
# the information-gap account curiosity is low when almost nothing is
# known (no question has formed) and low when almost everything is known
# (the gap has closed); here C(r) = 4 r (1 - r), largest at r = 0.5 - an
# assumed shape, in line with the inverted U measured by Kang et al.
# Curiosity fades after the teaser: the purchase hazard at time t (days)
# is h(t) = h0 C(r) exp(-t / tau), h0 = 1.2 per day, tau = 0.5 day.
# If the paywall pops up at delay T, the infinite-horizon cumulative
# purchase conversion is
#   1 - exp(-h0 C tau exp(-T / tau)).
import math

H0, TAU = 1.2, 0.5

def p_buy(r, delay):
    c = 4 * r * (1 - r)
    return 1 - math.exp(-H0 * c * TAU * math.exp(-delay / TAU))

if __name__ == "__main__":
    print("   revealed r   curiosity   buy: offer at once"
          "   after 6 h   after 1 day")
    for r in (0.1, 0.3, 0.5, 0.7, 0.9):
        print("      {:3.1f}         {:4.2f}          {:5.3f}"
              "            {:5.3f}       {:5.3f}".format(
                  r, 4 * r * (1 - r), p_buy(r, 0), p_buy(r, 0.25),
                  p_buy(r, 1.0)))
    half = TAU * math.log(2)
    print("\n   curiosity half-life: {:.2f} days".format(half))
    print("\nReading: a teaser that reveals too little or too much leaves")
    print("a small gap; the middle maximises the chance to buy. Because")
    print("curiosity decays, an offer made a day later converts at about a")
    print("sixth of an offer made at once. Both shapes are assumptions to be")
    print("measured, not facts about any audience.")
#    revealed r   curiosity   buy: offer at once   after 6 h   after 1 day
#       0.1         0.36          0.194            0.123       0.029
#       0.3         0.84          0.396            0.263       0.066
#       0.5         1.00          0.451            0.305       0.078
#       0.7         0.84          0.396            0.263       0.066
#       0.9         0.36          0.194            0.123       0.029
#
#    curiosity half-life: 0.35 days
#
# Reading: a teaser that reveals too little or too much leaves
# a small gap; the middle maximises the chance to buy. Because
# curiosity decays, an offer made a day later converts at about a
# sixth of an offer made at once. Both shapes are assumptions to be
# measured, not facts about any audience.

揭示一半时(r=0.5r = 0.5)立即提供付费,最终购买的概率为 0.451;揭示太少(0.1)或太多(0.9)都只有 0.194。提供付费的时机同样重要:六小时后为 0.305,一天后只有 0.078,约为立即提供的六分之一。这两个形状——倒 U 与指数衰减——都是假设,需要在具体的受众中测量。

购买概率随免费钩子揭示比例 r 与付费提示延迟的变化。

「说得准」从何而来

口号的第二部分是「痛点直断 + 身份标签」:用户觉得说中了自己,就会转发。问题在于,觉得说中了,与真的说中了,是两回事。福勒在 1949 年给学生做了一份「个性测验」,然后发给每人同一份通用描述,学生们普遍给出接近满分的准确度评分 (Forer 1949);此后大量研究重复了这一「巴纳姆效应」(Dickson and Kelly 1985)。信号检测论把这个问题说得更精确:一个判断有没有信息,要看它能否把「真」与「假」区分开,而不是看它被认为对的比例 (Green and Swets 1966)。

第八段程序模拟 40 条陈述,每条对随机一人为真的基础概率在 0.2 到 0.9 之间。通用解读对所有人都给出基础概率最高的 6 条;有信息的解读以 0.8 的准确度知道每个人的真实情况,并为这个人挑出 6 条。然后做卡尔森式的调包检验:把自己的解读与两份别人的解读放在一起,看自己的那份是否得分最高 (Carlson 1985)。

推导

# Chapter 4. "It describes me exactly": perceived accuracy without signal.
# 40 statements; statement j is true of a random person with base rate
# b_j, drawn uniform on [0.2, 0.9]. 5,000 people; truth is independent
# across statements and people. A reading consists of 6 statements.
#   generic    the same 6 statements for everyone: those with the highest
#              base rates (a "Barnum" reading; no information about anyone)
#   informed   a reader who knows each truth value with accuracy 0.8 and
#              picks 6 statements it believes true of THIS person
# Perceived accuracy = share of the 6 statements true of the person.
# Swap test (as in Carlson's double-blind test): shown their own reading
# and two readings made for others, how often does a person's own reading
# score highest (ties split)? Chance is 1/3.
# Sharing is modelled as P(share) = 1 / (1 + exp(3 - 4 acc - 0.8 tag)),
# tag = 1 if the reading carries an identity label ("you are the kind of
# person who..."); the coefficients are assumptions.
import math
import random

def make(rng):
    base = [rng.uniform(0.2, 0.9) for _ in range(40)]
    people = [[rng.random() < b for b in base] for _ in range(5000)]
    return base, people

def generic(base):
    top = sorted(range(40), key=lambda j: -base[j])[:6]
    return lambda person, rng: top

def informed(base):
    def pick(person, rng):
        belief = [t if rng.random() < 0.8 else (not t) for t in person]
        cand = [j for j in range(40) if belief[j]]
        rng.shuffle(cand)
        cand.sort(key=lambda j: -base[j])
        return (cand + list(range(40)))[:6]
    return pick

def acc(person, stmts):
    return sum(person[j] for j in stmts) / len(stmts)

def share(a, tag):
    return 1 / (1 + math.exp(3 - 4 * a - 0.8 * tag))

if __name__ == "__main__":
    rng = random.Random(63008)
    base, people = make(rng)
    print("   reader     accuracy own   own wins swap test   share no tag"
          "   share tag")
    for name, maker in (("generic", generic), ("informed", informed)):
        pick = maker(base)
        readings = [pick(p, rng) for p in people]
        own = [acc(p, r) for p, r in zip(people, readings)]
        wins = 0.0
        for i, p in enumerate(people):
            others = [readings[(i + 1) % 5000], readings[(i + 2) % 5000]]
            scores = [own[i]] + [acc(p, o) for o in others]
            best = max(scores)
            if scores[0] == best:
                wins += 1 / scores.count(best)
        m = sum(own) / len(own)
        print("   {:8s}      {:5.3f}             {:5.3f}"
              "            {:5.3f}        {:5.3f}".format(
                  name, m, wins / len(people),
                  sum(share(a, 0) for a in own) / len(own),
                  sum(share(a, 1) for a in own) / len(own)))
    print("\nReading: the generic reading feels almost as accurate as the")
    print("informed one - both pick statements with high base rates - yet")
    print("it fails the swap test exactly at chance (1/3). Adding an")
    print("identity label raises sharing more than real information does.")
    print("Share rates measure how a reading feels, not whether it is right.")
#    reader     accuracy own   own wins swap test   share no tag   share tag
#    generic       0.838             0.333            0.583        0.745
#    informed      0.939             0.653            0.676        0.819
#
# Reading: the generic reading feels almost as accurate as the
# informed one - both pick statements with high base rates - yet
# it fails the swap test exactly at chance (1/3). Adding an
# identity label raises sharing more than real information does.
# Share rates measure how a reading feels, not whether it is right.

通用解读被认为的准确度是 0.838,与有信息的解读(0.939)相差不大——两者都挑了基础概率高的陈述。但通用解读在调包检验中赢的比例正好是随机水平 0.333:它对每个人都一样「准」,所以无法区分任何人。有信息的解读赢了 0.653。

更值得注意的是转发。在假设的转发模型里,给通用解读加上身份标签,转发率从 0.583 升到 0.745,超过了不加标签的有信息解读(0.676)。转发率衡量的是一份解读给人的感觉,不是它对不对;用转发率来优化解读,会把系统推向巴纳姆式的通用陈述。

假说 · 尚未验证

作者的假说(尚未验证): 在一个真实的测算网站上,按转发率排序的解读模板,其调包检验得分与随机水平没有可区分的差别。什么结果会证明它错:在预先登记的盲法调包检验中,转发率最高的模板让用户选中自己解读的比例显著高于三分之一。

小结

信息缺口理论把好奇心写成一个先升后降、随时间衰减的量:免费钩子揭示一半时购买概率最高,延迟一天提供付费,概率降到约六分之一。「说得准」的感觉不需要任何信息:通用解读被认为的准确度接近有信息的解读,却在调包检验中只有随机水平;身份标签提高转发的幅度大于真实信息。转发率衡量的是感觉,不是准确性。

参考文献

Carlson, Shawn. 1985. “A Double-Blind Test of Astrology.” Nature 318: 419–25.
Dickson, D. H., and I. W. Kelly. 1985. “The ’Barnum Effect’ in Personality Assessment: A Review of the Literature.” Psychological Reports 57 (2): 367–82.
Forer, Bertram R. 1949. “The Fallacy of Personal Validation: A Classroom Demonstration of Gullibility.” Journal of Abnormal and Social Psychology 44 (1): 118–23.
Golman, Russell, and George Loewenstein. 2018. “Information Gaps: A Theory of Preferences Regarding the Presence and Absence of Information.” Decision 5 (3): 143–64.
Green, David M., and John A. Swets. 1966. Signal Detection Theory and Psychophysics. Wiley.
Kang, Min Jeong, Ming Hsu, Ian M. Krajbich, et al. 2009. “The Wick in the Candle of Learning: Epistemic Curiosity Activates Reward Circuitry and Enhances Memory.” Psychological Science 20 (8): 963–73.
Loewenstein, George. 1994. “The Psychology of Curiosity: A Review and Reinterpretation.” Psychological Bulletin 116 (1): 75–98.

Fan, Y. (2026). Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization: Galton-Watson Referral Thresholds, Share-to-Unlock Incentives and Second-Degree Price Discrimination [超临界病毒级联、信息缺口风险率与多层付费墙优化] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090174 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).