关于「持续同调:Rips 过滤、并查集零维条码与模 2 约化一维条码」,《社群拓扑体检》给出的定量结论是什么?

结论:第二章用并查集求零维条码、用模 2 列约化求一维条码,在两百个合成社群上,点赞总数的 AUC 只有 0.427,一维条码前三长之和 L3 为 0.975;补量社群的真实核心增至 30 人时降到 0.673,此时须靠零维短条。

数据口径:书中所有社群与成员都是固定种子的合成数据,不使用任何真实社群或个人的数据;全部计算由书中程序生成,参数为示意;本书是非个人化经营建议,「水军识别」只产出供人工复核的嫌疑名单,不自动处罚任何成员。 第二章用并查集求零维条码、用模 2 列约化求一维条码,在两百个合成社群上,点赞总数的 AUC 只有 0.427,一维条码前三长之和 L3 为 0.975;补量社群的真实核心增至 30 人时降到 0.673,此时须靠零维短条。

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: 0.427, 0.673, 0.975, 30

metric AUC95 % interval
likes (activity)0.427 [0.348, 0.506]
co-engagement triangles0.003 [0.000, 0.011]
longest H1 bar0.970 [0.946, 0.994]
H1 lead-3 persistence L30.975 [0.953, 0.997]
short H0 bars per account0.000 [0.000, 0.000]
dose scan: real core of padded communities, 40 vs 40 each
core AUC likes AUC L3 AUC short-bars
100.570 0.998 0.000
150.187 0.963 0.000
200.000 0.918 0.000
250.000 0.786 0.000
300.000 0.673 0.000
查看完整验算程序
# Chapter 2. p23: Which number separates genuine from padded communities?
# 100 genuine (24-36 members over the whole ring) against 100 padded
# (10-14 real members over the whole ring plus 14-22 lockstep
# accounts), fixed seed, synthetic only. AUC = P(genuine scores above
# padded); 0.5 is no information, values near 0 mean the metric ranks
# padded communities higher. 95 % intervals by Hanley and McNeil (1982).
# Then a dose scan: the real core of the padded communities grows.
import math
import random

T, P = 24, 5          # 24 topics on a ring, 5 posts each (illustrative)
N_POSTS = T * P


def member(rng, centre, width, q=0.8, stray=0.03):
    # a genuine member likes posts on an arc of topics around centre
    s = set()
    for t in range(T):
        dd = abs(t - centre) % T
        dd = min(dd, T - dd)
        hit = q if dd <= width / 2.0 else stray
        for k in range(P):
            if rng.random() < hit:
                s.add(t * P + k)
    return s


def sock(rng, target, q=0.95, stray=0.01):
    # a boosting account likes the operator's target posts in lockstep
    s = set()
    for p in range(N_POSTS):
        if rng.random() < (q if p in target else stray):
            s.add(p)
    return s


def community(rng, n_real, arc, n_sock):
    # n_real genuine members with interest centres on [0, arc) topics,
    # plus n_sock boosting accounts aimed at 30 target posts
    pts = []
    for _ in range(n_real):
        pts.append(member(rng, rng.uniform(0, arc), rng.randint(5, 8)))
    target = set(rng.sample(range(N_POSTS), 30))
    for _ in range(n_sock):
        pts.append(sock(rng, target))
    return pts


def jaccard(a, b):
    u = len(a | b)
    return 1.0 - len(a & b) / u if u else 0.0


def barcodes(pts, cap=1.0):
    return barcodes_d(len(pts), lambda i, j: jaccard(pts[i], pts[j]), cap)


def barcodes_d(n, dist, cap=1.0):
    # Vietoris-Rips filtration (up to triangles) of a distance dist(i, j).
    # H0 by union-find; H1 by mod-2 reduction of triangle columns,
    # each column stored as a Python integer used as a bitset of edges.
    edges = []
    for i in range(n):
        for j in range(i + 1, n):
            d = dist(i, j)
            if d <= cap:
                edges.append((d, i, j))
    edges.sort()
    eid = {(i, j): k for k, (d, i, j) in enumerate(edges)}
    par = list(range(n))

    def find(x):
        while par[x] != x:
            par[x] = par[par[x]]
            x = par[x]
        return x
    h0, neg = [], set()
    for k, (d, i, j) in enumerate(edges):
        a, b = find(i), find(j)
        if a != b:
            par[max(a, b)] = min(a, b)
            h0.append((d, i, j))
            neg.add(k)
    tris = []
    for (i, j), e in eid.items():
        for k in range(j + 1, n):
            f, g = eid.get((i, k)), eid.get((j, k))
            if f is not None and g is not None:
                tris.append(tuple(sorted((e, f, g), reverse=True)))
    tris.sort()
    piv, h1 = {}, []
    for t in tris:
        col = (1 << t[0]) | (1 << t[1]) | (1 << t[2])
        while col:
            low = col.bit_length() - 1
            if low not in piv:
                piv[low] = col
                break
            col ^= piv[low]
        if col and edges[t[0]][0] > edges[low][0]:
            h1.append((edges[low][0], edges[t[0]][0]))
    for k in range(len(edges)):
        if k not in neg and k not in piv:
            h1.append((edges[k][0], cap))     # still open at the cap
    return h0, h1


def lead3(h1):
    # length of the three longest one-dimensional bars, added exactly
    top = sorted((d - b for b, d in h1), reverse=True)[:3]
    return math.fsum(top)


def short_frac(h0, n, delta=0.2):
    # zero-dimensional bars dying below delta, per account
    return len([1 for d, i, j in h0 if d < delta]) / float(n)


def auc(pos, neg):
    # Mann-Whitney estimate of P(score of pos > score of neg)
    wins = 0
    for a in pos:
        for b in neg:
            wins += 2 if a > b else (1 if a == b else 0)
    return wins / (2.0 * len(pos) * len(neg))


def auc_se(a, n1, n2):
    # Hanley and McNeil (1982) standard error of an AUC
    q1, q2 = a / (2.0 - a), 2.0 * a * a / (1.0 + a)
    v = (a * (1 - a) + (n1 - 1) * (q1 - a * a)
         + (n2 - 1) * (q2 - a * a)) / (n1 * n2)
    return math.sqrt(max(v, 0.0))


def features(pts):
    n = len(pts)
    likes = 0
    for p in pts:
        likes += len(p)
    tri = 0
    near = [[jaccard(pts[i], pts[j]) <= 0.75 for j in range(n)]
            for i in range(n)]
    for i in range(n):
        for j in range(i + 1, n):
            if near[i][j]:
                for k in range(j + 1, n):
                    if near[i][k] and near[j][k]:
                        tri += 1
    h0, h1 = barcodes(pts)
    longest = max([d - b for b, d in h1] + [0.0])
    return [likes, tri, longest, lead3(h1), short_frac(h0, n)]


rng = random.Random(20261113)
G = [features(community(rng, rng.randint(24, 36), 24, 0))
     for _ in range(100)]
F = [features(community(rng, rng.randint(10, 14), 24,
                        rng.randint(14, 22))) for _ in range(100)]
names = ["likes (activity)", "co-engagement triangles",
         "longest H1 bar", "H1 lead-3 persistence L3",
         "short H0 bars per account"]
print("%-27s %6s  %15s" % ("metric", "AUC", "95 % interval"))
for m, name in enumerate(names):
    a = auc([g[m] for g in G], [f[m] for f in F])
    se = auc_se(max(a, 1 - a), 100, 100)
    print("%-27s %6.3f  [%.3f, %.3f]" % (name, a, max(0.0, a - 1.96 * se),
                                         min(1.0, a + 1.96 * se)))
print("dose scan: real core of padded communities, 40 vs 40 each")
print("core  AUC likes  AUC L3  AUC short-bars")
G40 = G[:40]
for core in (10, 15, 20, 25, 30):
    PAD = [features(community(rng, core, 24, 18)) for _ in range(40)]
    print("%4d  %9.3f  %6.3f  %14.3f" % (
        core, auc([g[0] for g in G40], [p[0] for p in PAD]),
        auc([g[3] for g in G40], [p[3] for p in PAD]),
        auc([g[4] for g in G40], [p[4] for p in PAD])))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《持续同调拓扑贝蒂数、高净值私域社群结构空洞与虚假抱团水军识别母典》第 2 章「持续同调:Rips 过滤、并查集零维条码与模 2 约化一维条码」(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-236 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026236,
  author    = {Fan, Yafu},
  title     = {Persistent Homology and Betti Numbers: Structural Voids in High-Net-Worth Private Communities and the Review of Fake Cliques},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论