关于「书库主题聚类与合集重叠」,《蓝海词与主题合集》给出的定量结论是什么?
结论:用关键词集合的杰卡德相似度与平均连接法,可以把书库自动聚成主题,示意书库里类内相似度 0.269,类间只有 0.007。主题合集的目标读者平均还有六成以上的价值没拿到,比跨主题合集更值得推;可是将近四分之一的目标读者已买过合集的一半以上,原价推销对他们几乎无效。
数据口径:对小作者,关键词带来的期望销量约等于搜索需求、进入前几位的概率、点击率与转化率之积;大词的进前位概率近乎为零,示意数据里周搜索两万次的大词只带来约 0.0254 本周销量,冷僻短语却有 3.1445 本。 核对原文后,本书做了蓝图修正:排名换销量的指数是示意参数而非定值;KDP 关键词规则以官方帮助页为准,每本书至多七个关键词,不得用与本书无关的作者名或自己不拥有的品牌;AI 生成内容须如实披露;本书不引用艾瑞里的任何研究。
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: 0.007, 0.269
| volumes 24, clusters 9 at cut 0.12 | |
| V01 V02 V03 V04 V05 | pricing, bundling, menu |
| V06 V07 V08 V10 | ranking, search, demand |
| V09 | crawl, indexing, long-tail |
| V11 V12 | routing, time-windows, buyers |
| V13 V14 | inventory, lead-time, restock |
| V15 V16 V17 | amm, arbitrage, liquidity |
| V18 V19 V20 | property, mortgage, rent |
| V21 V22 | cashflow, delivery, reliability |
| V23 V24 | abstracts, catalogue, library |
| mean Jaccard inside clusters 0.269, across 0.007 | |
| pairs compared: 276 |
查看完整验算程序
# Chapter 4. p41: clustering a catalogue into themes by Jaccard overlap.
# Synthetic catalogue of 24 volumes (ids V01..V24), each tagged with a
# keyword set; tags are illustrative, loosely modelled on the kinds of
# subjects in a bilingual monograph library. Similarity of two volumes
# is |A & B| / |A | B| (Jaccard). Average-linkage agglomerative
# clustering merges the closest pair until no pair of clusters has
# mean similarity above CUT. Ties break on the smallest ids, so the
# result does not depend on hash order.
import itertools
import math
CUT = 0.12
CAT = {
"V01": "bundling pricing information-goods correlation revenue",
"V02": "bundling mixed-bundling pricing screening menu",
"V03": "pricing left-digit price-ending perception menu",
"V04": "bundling licensing tiers pricing revenue",
"V05": "screening menu incentive tiers pricing",
"V06": "keyword search zipf demand ranking",
"V07": "keyword long-tail ranking sales-rank demand",
"V08": "search ranking click position cascade",
"V09": "seo long-tail indexing search crawl",
"V10": "generative-search citation ranking search",
"V11": "routing tsp time-windows buyers shopping",
"V12": "routing vehicle depot time-windows queue",
"V13": "queue inventory restock buyers lead-time",
"V14": "inventory restock newsvendor demand lead-time",
"V15": "liquidity amm lvr volatility arbitrage",
"V16": "liquidity vault kelly tail-risk volatility",
"V17": "arbitrage sandwich slippage amm mev",
"V18": "property valuation rent yield mortgage",
"V19": "property mortgage amortisation yield cashflow",
"V20": "rent valuation hedonic property location",
"V21": "cashflow refund chargeback delivery reliability",
"V22": "reliability delivery queue poisson cashflow",
"V23": "catalogue classification library abstracts",
"V24": "library catalogue abstracts bundling access",
}
def jac(a, b):
return len(a & b) / len(a | b)
def cluster(sets, cut):
ids = sorted(sets)
sim = {(a, b): jac(sets[a], sets[b])
for a, b in itertools.combinations(ids, 2)}
groups = [[i] for i in ids]
def link(g, h):
vals = [sim[tuple(sorted((a, b)))] for a in g for b in h]
return math.fsum(vals) / len(vals)
while len(groups) > 1:
best = None
for x, y in itertools.combinations(range(len(groups)), 2):
s = link(groups[x], groups[y])
if best is None or s > best[0] + 1e-12:
best = (s, x, y)
if best[0] <= cut:
break
_, x, y = best
groups[x] = sorted(groups[x] + groups[y])
del groups[y]
return sorted(groups, key=lambda g: g[0]), sim
if __name__ == "__main__":
sets = {k: set(v.split()) for k, v in CAT.items()}
groups, sim = cluster(sets, CUT)
inside, across = [], []
label = {i: n for n, g in enumerate(groups) for i in g}
for (a, b), s in sorted(sim.items()):
(inside if label[a] == label[b] else across).append(s)
print("volumes {}, clusters {} at cut {:.2f}".format(
len(sets), len(groups), CUT))
for g in groups:
tags = {}
for i in g:
for t in sets[i]:
tags[t] = tags.get(t, 0) + 1
top = sorted(tags, key=lambda t: (-tags[t], t))[:3]
print(" {:<28} {}".format(" ".join(g), ", ".join(top)))
print("mean Jaccard inside clusters {:.3f}, across {:.3f}".format(
math.fsum(inside) / len(inside), math.fsum(across) / len(across)))
print("pairs compared: {:,}".format(len(sim)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《百卷数字专著蓝海词卡位、高客单主题合集直售与全自动知识收银母典》第 4 章「书库主题聚类与合集重叠」(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-356 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026356,
author = {Fan, Yafu},
title = {Niche Keyword Positioning for Digital Monograph Catalogues, High-Ticket Thematic Bundle Direct Sales and Automated Knowledge Cashflow},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}