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
第五章 同意的代价与推荐的数量
跨站点使用个人数据需要同意
第三章的路由用的是读者在一个站点上的行为,去决定另一个站点向他推荐什么。这是对个人数据的使用;按马来西亚《2010 年个人数据保护法》,处理个人数据须告知用途,用于直接营销须取得同意,用户也可以要求停止 (Parliament of Malaysia 2010)。设只有比例 的读者同意,其余读者只能看到按当前页面内容(不涉及个人数据)给出的推荐,点击率较低。
推导
# Chapter 5. p09: Cross-site routing needs consent - what it costs.
# Routing a user from site i to site j by their history on i uses
# personal data across sites; under personal-data law the user must be
# told and, for marketing use, consent. If a share c consents, only they
# can be routed by history; the rest see a context-only recommendation
# (based on the page they are on, no personal data) with a lower
# click-through. Revenue per 1,000 report readers (illustrative).
def revenue(c, ctr_hist=0.040, ctr_ctx=0.025, conv=0.08, value=98.0):
clicks = 1000 * (c * ctr_hist + (1 - c) * ctr_ctx)
return clicks * conv * value
if __name__ == "__main__":
full = revenue(1.0)
print(" consent share revenue per 1,000 readers vs all consenting")
for c in (0.0, 0.3, 0.5, 0.7, 1.0):
r = revenue(c)
print(" {:13.1f} {:24.2f} {:.3f}".format(c, r, r / full))
print(" context-only routing keeps {:.3f} of the history-based"
" revenue".format(revenue(0.0) / full))
# consent share revenue per 1,000 readers vs all consenting
# 0.0 196.00 0.625
# 0.3 231.28 0.737
# 0.5 254.80 0.812
# 0.7 278.32 0.887
# 1.0 313.60 1.000
# context-only routing keeps 0.625 of the history-based revenue示意参数下,全部同意时每 1,000 名读者带来 RM 313.60;一半同意时 RM 254.80,是全部同意的 0.812;没有人同意时,只按页面内容推荐也能保住 0.625。按内容推荐不需要个人数据,是一个不依赖同意的下限;同意带来的是这个下限之上的部分。
注
设计原则。 默认使用按内容推荐;只在读者明确同意后启用按历史的跨站点路由;同意与撤回同样容易。第 77 本第五章的结论在这里同样成立:被发现的暗中做法,其代价远大于多出的那几成点击。
报告底部放几条推荐
推导
# Chapter 5. p10: How many cross-sell links at the foot of a report?
# Cascade model: the reader scans links top to bottom; at each link they
# click with probability a_k (relevance falls with rank: a_k = 0.06 x
# 0.7^(k-1)) and otherwise continue with probability s (attention left,
# s = 0.6). More links also dilute: each extra link lowers s by 0.05.
# Expected clicks for k links.
def clicks(k, a0=0.06, decay=0.7, s0=0.6, dilution=0.05):
s = max(0.0, s0 - dilution * (k - 1))
reach, total = 1.0, 0.0
for i in range(k):
a = a0 * decay ** i
total += reach * a
reach *= (1 - a) * s
return total
if __name__ == "__main__":
for k in range(1, 7):
print(" links {}: expected clicks per reader {:.4f}".format(
k, clicks(k)))
best = max(range(1, 7), key=clicks)
print(" best number of links: {}".format(best))
# links 1: expected clicks per reader 0.0600
# links 2: expected clicks per reader 0.0817
# links 3: expected clicks per reader 0.0864
# links 4: expected clicks per reader 0.0848
# links 5: expected clicks per reader 0.0815
# links 6: expected clicks per reader 0.0781
# best number of links: 3用级联模型描述读者从上到下浏览推荐:每一条以随排名递减的概率被点击,否则以一定概率继续往下看;每多一条推荐,继续浏览的概率还会被稀释。示意参数下,1 条推荐的期望点击为 0.0600,2 条 0.0817,3 条 0.0864,4 条起开始下降;最优是 3 条,第 3 条只增加约 6%。「Top-2」是一个合理的简化,但最优数量由点击率的递减速度与稀释程度决定,需要测量。
小结
跨站点按历史推荐需要同意;在示意参数下,只按页面内容推荐也能保住约六成收入。报告底部的推荐条数存在最优值,示意参数下为 3 条,第 2、3 条之间的差别不大。
参考文献
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).