关于「破冰配对的最大权匹配:精确递推、贪心、轮转与三个可能世界」,《私董社群的互惠边与破冰配对》给出的定量结论是什么?

结论:第四章用位掩码递推求精确最大权匹配,在三个可能世界中检验它对随机配对的优势:评分正确时每轮多建 0.766 条互惠边,属性无关时为零,互补相吸时少建 0.552 条。

数据口径:书中的社群、成员、属性与概率均为示意;全部计算由书中程序生成;本书是非个人化经营建议,只使用成员明示同意的数据,不做个人画像,退群自由,退款按约定。 第一章普查一个示意社群:74 个有往来的有向对只形成 15 条互惠边,并用两个模拟世界说明,风险比 6.09 与 2.69 都可能出自互惠边毫无作用的世界。

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: 0.766

boards 200, mean zero-edge members per board 6.94
truthnew edges: plannedrandomgainzero-edge: plannedrandom
T13.1252.359+0.7662.3161.499
T23.2003.200+0.0002.4402.159
T31.7902.342-0.5521.3101.489

核心公式

\[\text{logit}\, p_{ij} = -1.6 + 0.5\,[\text{同行业}] + 0.4\,[\text{同城}] + 0.3\,[\text{规模档相差不超过一档}] + 0.12 \min(d_i + d_j, 6)\]
查看完整验算程序
# Chapter 4. p42: the pairing gain depends on which world is true.
# 200 illustrative boards of 16 (generator of p11, seeds 1000 to 1199).
# The planner always pairs by the score of p41 (exact optimum with the
# zero-edge priority). The pairing is then scored under three truths:
#   T1 the planner's score is right;
#   T2 attributes do not matter: every open pair has p = 0.40;
#   T3 complements attract: logit p = -0.9 - 0.5 [same sector]
#      - 0.3 [same city] + 0.12 min(d_i + d_j, 6).
# Random pairing = mean of 300 random valid pairings per board.
# Reported: new reciprocal edges per round, and zero-edge members who
# gain one, both as expectations, averaged over boards.
import math
import random
from functools import lru_cache

SECT = ["MFG", "RET", "SVC", "TEC"]
CITY = ["SH", "SZ", "BJ"]


def make_board(seed=278, n=16):
    rng = random.Random(seed)
    mem = []
    for _ in range(n):
        mem.append({"sec": SECT[rng.randrange(4)],
                    "city": CITY[rng.randrange(3)],
                    "band": 1 + rng.randrange(3),
                    "act": 0.15 + 0.85 * rng.random() ** 1.6})
    cnt = [[0] * n for _ in range(n)]
    for a in range(n):
        for b in range(n):
            if a == b:
                continue
            sim = ((mem[a]["sec"] == mem[b]["sec"])
                   + (mem[a]["city"] == mem[b]["city"])
                   + (abs(mem[a]["band"] - mem[b]["band"]) <= 1))
            q = mem[a]["act"] * (0.035 + 0.045 * sim)
            cnt[a][b] = len([t for t in range(6) if rng.random() < q])
    return mem, cnt


def logistic(z):
    return 1 / (1 + math.exp(-z))


def truth(k, mi, mj, di, dj):
    if k == 0:
        return logistic(-1.6 + 0.5 * (mi["sec"] == mj["sec"])
                        + 0.4 * (mi["city"] == mj["city"])
                        + 0.3 * (abs(mi["band"] - mj["band"]) <= 1)
                        + 0.12 * min(di + dj, 6))
    if k == 1:
        return 0.40
    return logistic(-0.9 - 0.5 * (mi["sec"] == mj["sec"])
                    - 0.3 * (mi["city"] == mj["city"])
                    + 0.12 * min(di + dj, 6))


def plan(n, rec, W):
    @lru_cache(maxsize=None)
    def best(mask):
        if mask == 0:
            return 0.0, ()
        i = (mask & -mask).bit_length() - 1
        top = (-1.0, ())
        for j in range(i + 1, n):
            if mask >> j & 1 and not rec[i][j]:
                v, pr = best(mask & ~(1 << i) & ~(1 << j))
                if v >= 0 and v + W[i][j] > top[0]:
                    top = (math.fsum([v, W[i][j]]), ((i, j),) + pr)
        return top
    return best((1 << n) - 1)[1]


rng = random.Random(20261113)
res = {(t, m): [] for t in range(3) for m in ("alg", "rnd")}
zres = {(t, m): [] for t in range(3) for m in ("alg", "rnd")}
nzero = []
for seed in range(1000, 1200):
    mem, cnt = make_board(seed)
    n = len(cnt)
    rec = [[bool(i != j and cnt[i][j] and cnt[j][i]) for j in range(n)]
           for i in range(n)]
    deg = [len([j for j in range(n) if rec[i][j]]) for i in range(n)]
    zero = {i for i in range(n) if deg[i] == 0}
    nzero.append(len(zero))
    T = [[[truth(k, mem[i], mem[j], deg[i], deg[j]) for j in range(n)]
          for i in range(n)] for k in range(3)]
    W = [[T[0][i][j] * (2 if i in zero or j in zero else 1)
          for j in range(n)] for i in range(n)]
    alg = plan(n, rec, W)
    rnd = []
    while len(rnd) < 300:
        o = list(range(n))
        rng.shuffle(o)
        prs = [(o[k], o[k + 1]) for k in range(0, n, 2)]
        if all(not rec[i][j] for i, j in prs):
            rnd.append(prs)
    for t in range(3):
        res[(t, "alg")].append(math.fsum(T[t][i][j] for i, j in alg))
        zres[(t, "alg")].append(math.fsum(
            T[t][i][j] for i, j in alg if i in zero or j in zero))
        e = [math.fsum(T[t][i][j] for i, j in prs) for prs in rnd]
        ez = [math.fsum(T[t][i][j] for i, j in prs
                        if i in zero or j in zero) for prs in rnd]
        res[(t, "rnd")].append(math.fsum(e) / len(e))
        zres[(t, "rnd")].append(math.fsum(ez) / len(ez))
m = len(nzero)
print("boards %d, mean zero-edge members per board %.2f"
      % (m, math.fsum(nzero) / m))
print("truth   new edges: planned  random  gain    zero-edge: planned"
      "  random")
for t, name in enumerate(("T1", "T2", "T3")):
    a = math.fsum(res[(t, "alg")]) / m
    r = math.fsum(res[(t, "rnd")]) / m
    za = math.fsum(zres[(t, "alg")]) / m
    zr = math.fsum(zres[(t, "rnd")]) / m
    print("  %s               %.3f   %.3f  %+.3f              %.3f   %.3f"
          % (name, a, r, a - r, za, zr))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《高客单私董社群防退群内聚拓扑、度中心性互惠边权重与破冰配对算法母典》第 4 章「破冰配对的最大权匹配:精确递推、贪心、轮转与三个可能世界」(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-278 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026278,
  author    = {Fan, Yafu},
  title     = {Cohesion Topology of High-Ticket Peer Boards and First-Quarter Silent Attrition: Degree Centrality, Reciprocal Edge Weights and Ice-Breaking Pairing Algorithms},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论