关于「五张图,五个目标:兴趣图、搜索图、关注图、对话图与熟人图」,《五大平台算法与一鱼五吃》给出的定量结论是什么?
结论:第一章说明五个平台从五张不同的图挑候选,五种示意帖型没有一种通吃,按平台挑帖型的总分是只用一种的 1.496 倍;X 在 2023 年公开的权重里回复是点赞的 27 倍、作者回应的回复是 150 倍,但权重乘的是每位读者的预测概率,不是累计次数。
数据口径:书中把每条平台说法标为三类:带日期的官方披露、同行评审研究、本书的示意模型;书中的权重、概率与漏斗参数均为示意,不是任何平台的真实设定,全部计算由书中程序生成。 本书是一般性经营参考,不构成法律、税务或投资建议;不涉及假账号、机器人、互赞群、买互动、抓取个人资料或群发私讯;合作内容一律标明;达人酬劳只有单层直付,不涉及任何多层分润。
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: 1.496
fit = score / score of the best archetype on that platform (1.00 = best)
RedNote Threads X Instagram Facebook
search guide 1.00 0.21 -0.16 0.75 0.47
open question 0.32 1.00 1.00 0.44 0.59
hot take -0.01 0.24 -0.21 -0.00 -0.15
visual set 0.59 0.39 -0.06 1.00 0.75
local notice 0.29 0.38 0.07 0.47 1.00
best on RedNote : search guide
best on Threads : open question
best on X : open question
best on Instagram: visual set
best on Facebook : local notice
archetypes that win somewhere: 4 of 5
one archetype on all five (open question): total fit 3.34 of 5.00
tailored archetype per platform: total fit 5.00, gain x 1.496查看完整验算程序
# Chapter 1. p11: four post archetypes scored by five parameterised ranking functions.
# Every platform is modelled as score = fsum(weight[a] * prob[a]) over actions a.
# The X row borrows four weights that X published on 5 April 2023 for its heavy
# ranker (fav 0.5, retweet 1.0, reply 13.5, reply engaged by author 75.0,
# negative feedback -74.0); every other weight and every probability below is
# illustrative, invented for this book, and is NOT any platform's real setting.
import math
ACTIONS = ("dwell", "like", "save", "reply", "areply", "share", "dm", "search", "group", "neg")
WEIGHTS = { # illustrative except the five X values named above
"RedNote": (1.0, 1.0, 4.0, 2.0, 2.0, 3.0, 0.0, 6.0, 0.0, -20.0),
"Threads": (0.5, 1.0, 0.5, 12.0, 25.0, 3.0, 1.0, 0.0, 0.0, -30.0),
"X": (0.0, 0.5, 0.0, 13.5, 75.0, 1.0, 0.0, 0.0, 0.0, -74.0),
"Instagram": (2.0, 1.0, 3.0, 2.0, 2.0, 6.0, 8.0, 0.5, 0.0, -30.0),
"Facebook": (1.0, 1.0, 0.5, 4.0, 4.0, 5.0, 2.0, 0.0, 8.0, -30.0),
}
POSTS = { # illustrative probabilities per impression
"search guide": (0.30, 0.04, 0.060, 0.004, 0.001, 0.010, 0.005, 0.080, 0.002, 0.005),
"open question": (0.18, 0.03, 0.010, 0.035, 0.015, 0.008, 0.004, 0.004, 0.006, 0.004),
"hot take": (0.15, 0.05, 0.005, 0.030, 0.010, 0.020, 0.004, 0.005, 0.004, 0.020),
"visual set": (0.35, 0.08, 0.030, 0.006, 0.001, 0.015, 0.030, 0.010, 0.002, 0.004),
"local notice": (0.20, 0.03, 0.010, 0.015, 0.004, 0.025, 0.006, 0.004, 0.050, 0.006),
}
BASE = (0.25, 0.04, 0.020, 0.010, 0.003, 0.015, 0.008, 0.020, 0.010, 0.003)
def score(w, p):
return math.fsum(wi * pi for wi, pi in zip(w, p))
raw = {n: [score(w, p) for w in WEIGHTS.values()] for n, p in POSTS.items()}
top = [max(raw[n][j] for n in POSTS) for j in range(len(WEIGHTS))]
print("fit = score / score of the best archetype on that platform (1.00 = best)")
print("%-14s" % "" + "".join("%11s" % k for k in WEIGHTS))
fit = {}
for name in POSTS:
fit[name] = [raw[name][j] / top[j] for j in range(len(WEIGHTS))]
print("%-14s" % name + "".join("%11.2f" % f for f in fit[name]))
best = [max(POSTS, key=lambda n: fit[n][j]) for j in range(len(WEIGHTS))]
for plat, name in zip(WEIGHTS, best):
print("best on %-9s: %s" % (plat, name))
print("archetypes that win somewhere: %d of %d" % (len(set(best)), len(POSTS)))
one = max(POSTS, key=lambda n: math.fsum(fit[n]))
single = math.fsum(fit[one])
print("one archetype on all five (%s): total fit %.2f of 5.00" % (one, single))
print("tailored archetype per platform: total fit 5.00, gain x %.3f" % (5.0 / single))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《五大社交图谱算法异质性、跨平台同源异构分发与达人聚拢母典》第 1 章「五张图,五个目标:兴趣图、搜索图、关注图、对话图与熟人图」(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-359 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026359,
author = {Fan, Yafu},
title = {Algorithmic Heterogeneity across RedNote, Threads, X, Instagram and Facebook: Cross-Platform Adaptation and Creator Aggregation},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}