5D Library › 机制设计、维克里—克拉克—格罗夫斯(VCG)拍卖与数字社群激励相容工程 · English

第六章 社群激励:积分通胀、徽章与女巫攻击

前五章的机制都有明确的「物品」与「支付」。线上社群的激励规则——积分、等级、徽章、排行榜、声誉分——也可以用同一套语言分析:积分是一种只在社群内流通的货币,徽章是一种信号与目标,声誉分是一个把许多人的评价汇总成一个数的函数。本章讨论三个问题:积分为什么会「通胀」;徽章怎样改变行为、证据有多强;声誉函数在有人可以低成本注册多个身份时能否被攻破。本章用到的社群数据全部是模拟的;积分通胀模型与女巫成本边界是作者的模型。

积分是一种货币

一个社群若允许用积分兑换稀缺的东西(辅导时段、线下活动名额、周边商品),积分就有了价格:一单位奖励值多少积分。虚拟经济的设计文献把发放积分的规则称为「水龙头」,把消耗积分的规则称为「排水口」,并讨论了它们的平衡问题 (Lehdonvirta and Castronova 2014)。最直接的类比是费雪的货币数量方程 MV=PTMV = PT (Fisher 1911):流通中的积分存量 MM、每期被拿去兑换的比例 VV、每期可兑换的奖励数量 YY 与积分价格 PP 之间,在「按市场出清价格兑换」的假设下满足 P=VM/YP = VM/Y。这只是类比:社群积分通常不能自由交易,价格往往由管理者规定,规定价格之下的「通胀」表现为排队、抢不到和兑换门槛不断提高。

作者的模型。 每月成员数 NtN_t 以 4% 增长,每名成员获得 e=100e = 100 积分;每月有 Y=200Y = 200 单位奖励按出清价格兑换,存量的 ν=25%\nu = 25\% 被拿去竞拍,兑换后的积分被销毁;可选的积分过期率为 λ\lambda:

Mt+1=(1−ν−λ)Mt+eNt,Pt=νMt/Y.M_{t+1} = (1 – \nu – \lambda)\,M_t + e\,N_t, \qquad P_t = \nu M_t / Y .

五个情景:A 每人积分固定;B 发放总量固定(成员越多,每人越少);C 在 A 的基础上每月 4% 过期;D 奖励数量随成员数同比增加;E 在 A 的基础上,从第 12 个月起一部分成员用自动化手段刷取廉价行为,获得 5 倍积分(比例每月增加 5 个百分点,最多 30%)。

推导

# Chapter 6. The author's model of points inflation in an online community (an analogy
# with the quantity equation M V = P Y, not a model of any real platform; all numbers
# are illustrative). Monthly: members N_t grow at g; each member earns e points; the
# community offers Y reward units (e.g. mentoring slots) that are allocated at a
# market-clearing points price; a fraction nu of the outstanding stock M_t is bid each
# month, so P_t = nu M_t / Y, and redeemed points are burned:
#     M_{t+1} = (1 - nu) M_t + e_t N_t - lambda M_t      (lambda: expiry, a "sink")
# Scenarios: A fixed points per member; B fixed total issuance (per-member points fall
# as the community grows); C A plus 4 % monthly expiry; D A, rewards grow with members;
# E A, plus "farming": from month 12 a share of members automate cheap actions and
# earn 5 times the points.
def run(T=36, N0=1000, g=0.04, e=100.0, Y0=200.0, nu=0.25, mode="A"):
    N, M = float(N0), e * N0 / nu                  # start in the steady state of A
    path = []
    for t in range(T + 1):
        Y = Y0 * (N / N0 if mode == "D" else 1.0)
        P = nu * M / Y; path.append(P)
        if mode == "B":
            issue = e * N0
        elif mode == "E" and t >= 12:
            farm = min(0.3, 0.05 * (t - 11))       # farming spreads 5 points a month
            issue = e * N * (1 - farm) + 5 * e * N * farm
        else:
            issue = e * N
        lam = 0.04 if mode == "C" else 0.0
        M = (1 - nu - lam) * M + issue; N *= 1 + g
    return path

if __name__ == "__main__":
    labels = {"A": "fixed points per member", "B": "fixed total issuance",
              "C": "A + 4 % monthly expiry", "D": "rewards grow with members",
              "E": "A + point farming from month 12"}
    print("points price of one reward unit (month 0 = steady state of A, e N / Y = 500)")
    print("  scenario                           m0      m12      m24      m36  "
          "value of 100 pts")
    for k, lab in labels.items():
        p = run(mode=k)
        print(f"  {k} {lab:31s} {p[0]:6.1f} {p[12]:8.1f} {p[24]:8.1f} {p[36]:8.1f}"
              f"  {100 / p[36]:6.3f} units")
    a, b = run(mode="A"), run(mode="B")
    print(f"A: monthly inflation in the last year {(a[36] / a[24]) ** (1 / 12) - 1:.2%}"
          f" (membership growth 4 %); B: {(b[36] / b[24]) ** (1 / 12) - 1:+.2%}")
    print(f"B keeps prices stable, but points per member fall to"
          f" {100 / 1.04 ** 36:.1f} by month 36 (from 100)")
# points price of one reward unit (month 0 = steady state of A, e N / Y = 500)
#   scenario                           m0      m12      m24      m36  value of 100 pts
#   A fixed points per member          500.0    692.3   1104.9   1768.9   0.057 units
#   B fixed total issuance             500.0    500.0    500.0    500.0   0.200 units
#   C A + 4 % monthly expiry           500.0    608.4    971.0   1554.5   0.064 units
#   D rewards grow with members        500.0    432.4    431.1    431.0   0.232 units
#   E A + point farming from month 12  500.0    692.3   2361.7   3889.5   0.026 units
# A: monthly inflation in the last year 4.00% (membership growth 4 %); B: +0.00%
# B keeps prices stable, but points per member fall to 24.4 by month 36 (from 100)

起点是情景 A 的稳态价格 eN/Y=500eN/Y = 500。情景 A 中价格 36 个月后升到 1,768.9,最后一年的月通胀率为 4.00%,恰好等于成员增长率:发放随人数增长、奖励不增长,积分就按人数增长的速度贬值;第 1 个月挣的 100 积分到第 36 个月只能换 0.057 单位奖励。过期(C)只能把价格水平压低一些(1,554.5),不能改变增长率。B 使价格稳定在 500,代价是每人每月积分从 100 降到 24.4,老成员会看到「同样的贡献得到的积分越来越少」。D 是唯一既稳定又不稀释的情景,但它要求真实的奖励供给随社群同步增长。E 显示刷分的效果:价格在第 24 个月达到 2,361.7,36 个月时 3,889.5,100 积分只值 0.026 单位。

这个模型给出的不是预测,而是一条会计关系:积分的实际价值取决于发放的增长与真实奖励的增长之差。它也说明了为什么「积分通胀」与古德哈特定律(第 27 本)常常同时出现:当积分按某个可刷的行为发放时,刷分既增加了积分供给,又使积分与真实贡献脱钩。

注

外在奖励的挤出效应(第 24 本)。 第 24 本讨论过:预期的、取决于参与或完成或表现的有形奖励,会降低人们对本来有兴趣的任务的内在动机(128 项实验的元分析);意外的奖励与言语的正反馈则没有这种效应或有正效应。积分若可以兑换实物,它更接近「有形奖励」;若只是记录与认可,它更接近「反馈」。这一区分在设计积分时与通胀问题同样重要,这里不再重复第 24 本的证据。

徽章:证据与一个简化模型

证据

安德森、胡滕洛赫尔、克莱因伯格与莱斯科维克用 Stack Overflow 从 2008 年 7 月上线到 2010 年底的活动记录研究徽章 (Anderson et al. 2013)。他们的模型中,用户有偏好的行为分布,偏离它有成本,徽章在某类行为达到阈值时给予一次性价值;模型预测用户在接近阈值时把行为「拐向」徽章,越接近拐得越多,拿到之后回到原来的分布。数据与之相符:例如「Electorate」徽章(对问题投票达到 600 次,并满足与对答案投票的比例条件)之前,用户对问题的投票增加、对答案的投票减少,得到徽章后很快回到基线;用户在获得这类徽章前的几天里总活动量也上升。他们还从模型中推出:单个徽章使目标行为最多的阈值在一个内部位置,远离原点;两个徽章最好大致均匀分布、价值相近。

关于游戏化的整体证据:哈马里、科伊维斯托与萨尔萨回顾了 24 项实证研究,大多数报告了正面效果,但效果依赖情境与用户,方法上有不少局限 (Hamari et al. 2014)。赛勒与霍姆纳的元分析发现,游戏化对学习的认知结局 g=0.49g = 0.49(19 项研究、1,686 人)、动机结局 g=0.36g = 0.36、行为结局 g=0.25g = 0.25;但在研究设计严格的子集中,只有认知结局的效应仍然显著(g=0.42g = 0.42),动机与行为结局不再显著;原作者把三项效应都称为显著的小效应 (Sailer and Homner 2020)。

下面的程序是安德森等人模型的一个简化重写(只有两种行为,没有「离开网站」的行为选择),不是他们的代码,数字也不能与原文比较。用户每一步以概率 xx 做目标行为 A1,偏好的概率是 p=0.05p = 0.05,偏离的成本为 2(x−p)22(x-p)^2,每一步以 θ=0.99\theta = 0.99 的概率留在网站,在第 kk 次 A1 时获得价值 VV 的徽章。

推导

# Chapter 6. A simplified two-action version of the badge model of Anderson,
# Huttenlocher, Kleinberg and Leskovec (2013); our simplification, not their code or
# their numbers. Each step the user takes action A1 with probability x (otherwise
# something else); her preferred rate is p, deviating costs c * 2 (x - p)^2 per step,
# she stays on the site to the next step with probability theta, and a badge worth V is
# awarded at the k-th A1 action (after which she reverts to p). With a = A1 actions so
# far, U(a) = max_x [ -2c(x - p)^2 + theta (x U(a+1) + (1 - x) U(a)) ], U(k) = V.
P, THETA, C, V = 0.05, 0.99, 1.0, 5.0
XS = [i / 2000 for i in range(2001)]

def solve(k, V=V):
    U, X = [0.0] * (k + 1), [0.0] * k
    U[k] = V
    for a in range(k - 1, -1, -1):
        best = None
        for x in XS:
            val = (-2 * C * (x - P) ** 2 + THETA * x * U[a + 1]) / (1 - THETA * (1 - x))
            if best is None or val > best[0]:
                best = (val, x)
        U[a], X[a] = best
    return U, X

def a1_yield(k, V=V):                              # expected lifetime A1 actions
    _, X = solve(k, V); A = P / (1 - THETA)        # after the badge: rate p
    for a in range(k - 1, -1, -1):
        x = X[a]; A = (x + THETA * x * A) / (1 - THETA * (1 - x))
    return A - P / (1 - THETA)                     # extra over no badge

if __name__ == "__main__":
    U, X = solve(60)
    print("badge at k = 60: optimal A1 rate x(a) by A1 actions already taken a")
    print("  a:  " + "".join(f"{a:7d}" for a in (0, 10, 20, 30, 40, 50, 55, 59)))
    print("  x:  " + "".join(f"{X[a]:7.3f}" for a in (0, 10, 20, 30, 40, 50, 55, 59)))
    print(f"  preferred rate p = {P}; after the badge x returns to {P}")
    ks = list(range(5, 121, 5))
    print(f"extra lifetime A1 actions induced (no-badge baseline {P / (1 - THETA):.1f})")
    print("  badge value V   k=5    k=10   k=20   k=40   k=80   best k on grid (yield)")
    for v in (5.0, 20.0, 50.0):
        ys = [a1_yield(k, v) for k in ks]; kb = ks[ys.index(max(ys))]
        cells = "".join(f"{ys[ks.index(k)]:7.2f}" for k in (5, 10, 20, 40, 80))
        print(f"  {v:13.0f}" + cells + f"   {kb:4d} ({max(ys):.2f})")
# badge at k = 60: optimal A1 rate x(a) by A1 actions already taken a
#   a:        0     10     20     30     40     50     55     59
#   x:    0.050  0.050  0.051  0.053  0.068  0.105  0.134  0.159
#   preferred rate p = 0.05; after the badge x returns to 0.05
# extra lifetime A1 actions induced (no-badge baseline 5.0)
#   badge value V   k=5    k=10   k=20   k=40   k=80   best k on grid (yield)
#               5   2.70   3.99   3.15   0.29   0.00     15 (4.00)
#              20   3.79   6.72   9.69   4.53   0.01     20 (9.69)
#              50   4.24   7.94  13.51  15.73   0.41     35 (16.44)

阈值为 60 时,用户在前 30 次几乎不改变行为(xx 从 0.050 到 0.053),越接近阈值拐得越多,最后一步前达到 0.159,拿到之后回到 0.05:这是原文所说的「加速」形状。第二张表是一生中额外诱发的 A1 行为(无徽章时的基线是 5.0 次):徽章价值 5 时,阈值 15 最好(多 4.00 次),阈值 40 只多 0.29 次,阈值 80 几乎没有效果,因为用户预计自己等不到那一天;徽章价值 20 与 50 时,最好的阈值分别移到 20 与 35。阈值太低,徽章「来得太容易」,诱发的行为少;太高,用户放弃。最好的阈值取决于徽章价值与用户留存,这些都需要估计,不能凭直觉定。

左:简化徽章模型中,用户选择的目标行为比例随已完成次数的变化(三种阈值),点线为偏好比例。右:第八章实验在三种个体差异假设下的把握度曲线(正态近似),竖线为计划的每组 150 人。

声誉与女巫攻击

社群里的声誉分把许多人的评价汇总成一个数。第 35 本讨论过声誉规范(形象分、简单持守与严格评判),那里的前提是每个人只有一个身份。线上社群的身份成本很低,这一前提往往不成立。

杜瑟在 2002 年把「一个实体伪装成多个身份」称为女巫攻击,并论证:没有一个(逻辑上集中的)可信认证机构时,除非假定各实体资源完全相当并且彼此协调这类不现实的条件,系统总可能受到女巫攻击 (Douceur 2002)。弗里德曼与雷斯尼克从经济学角度分析了「廉价化名」:如果离开再以新身份回来几乎不花钱,坏名声就没有约束力,均衡中新来者必须先「交学费」(以最低的信任开始),这对诚实的新人是一种社会成本 (Friedman and Resnick 2001)。程与弗里德曼把问题形式化:一个声誉函数若在攻击者复制出任意多个身份并互相推荐时,攻击者的声誉都不会上升,就称为防女巫;他们证明,对称的声誉函数(对节点的名字不加区分,如 PageRank (Page et al. 1999))不能防女巫,而基于「从可信种子出发的流」的非对称函数可以 (Cheng and Friedman 2005)。在实践中,TrustRank 只从人工审核的可信页面出发传播信任 (Gyöngyi et al. 2004);SybilGuard 利用社交网络中诚实区域与女巫区域之间只有少数「攻击边」这一结构来限制女巫身份的数量 (Yu et al. 2006)。

下面的程序在 200 名诚实成员的推荐网络上(每人随机推荐 4 人,5 名为可信种子)放入一名攻击者:他推荐 4 名诚实成员,并被 2 名诚实成员推荐(攻击边)。然后他制造 kk 个假身份,每个假身份推荐他,他也推荐每个假身份。程序比较三个声誉函数:PageRank、只向种子传送的个性化 PageRank(类似 TrustRank)、以及从种子到该节点的最大流。

推导

# Chapter 6. Sybil attacks on reputation (illustrative graph, not a real platform).
# 200 honest members, each endorsing 4 random others; 5 honest members are trusted
# seeds. An attacker (node 200) endorses 4 honest members and is endorsed by 2 of them
# ("attack edges"). He then creates k fake identities; each endorses him and he
# endorses each. Three reputation functions:
#   PR   PageRank, damping 0.85, uniform teleport (symmetric: every node treated alike)
#   TR   personalised PageRank teleporting to the seeds only (TrustRank-like)
#   MF   max-flow from the seeds to the node, unit capacity per endorsement
# Cheng and Friedman (2005): symmetric functions cannot be sybilproof; flow-based ones
# can. Last part: the author's cost-boundary model for a top-10 bonus.
import collections, random

def graph(k, rng_seed=6):
    rng = random.Random(rng_seed); n = 200; out = collections.defaultdict(set)
    for i in range(n):
        while len(out[i]) < 4:
            j = rng.randrange(n)
            if j != i:
                out[i].add(j)
    A = n
    out[A] |= set(rng.sample(range(n), 4))
    for i in rng.sample(range(5, n), 2):
        out[i].add(A)
    for s in range(n + 1, n + 1 + k):
        out[s].add(A); out[A].add(s)
    return out, n + 1 + k, A

def pagerank(out, N, tele, it=80, d=0.85):
    r = [1.0 / N] * N
    for _ in range(it):
        nr = [(1 - d) * tele[i] for i in range(N)]
        for i in range(N):
            share = d * r[i] / len(out[i])
            for j in out[i]:
                nr[j] += share
        r = nr
    return r

def maxflow(out, N, seeds, t):
    cap = collections.defaultdict(int); adj = collections.defaultdict(set); S = N
    for i in range(N):
        for j in out[i]:
            cap[i, j] += 1; adj[i].add(j); adj[j].add(i)
    for s in seeds:
        cap[S, s] = 10 ** 6; adj[S].add(s); adj[s].add(S)
    flow = 0
    while True:
        prev = {S: None}; q = collections.deque([S])
        while q and t not in prev:
            u = q.popleft()
            for v in adj[u]:
                if v not in prev and cap[u, v] > 0:
                    prev[v] = u; q.append(v)
        if t not in prev:
            return flow
        v = t
        while prev[v] is not None:
            u = prev[v]; cap[u, v] -= 1; cap[v, u] += 1; v = u
        flow += 1

if __name__ == "__main__":
    seeds = range(5); rows = []
    print("   k   PR score  PR rank   TR score  TR rank   MF (max-flow from seeds)")
    for k in (0, 1, 2, 5, 10, 20, 50):
        out, N, A = graph(k)
        pr = pagerank(out, N, [1.0 / N] * N)
        tr = pagerank(out, N, [0.2 if i < 5 else 0.0 for i in range(N)])
        hon = list(range(200)) + [A]                # rank among real members only
        rk = lambda sc: 1 + sum(sc[i] > sc[A] for i in hon)
        mf = maxflow(out, N, seeds, A); rows.append((k, rk(pr)))
        print(f"  {k:2d}  {pr[A] * 1000:8.3f}  {rk(pr):7d}  {tr[A] * 1000:9.3f}"
              f"  {rk(tr):7d}   {mf:3d}")
    out, N, A = graph(0)
    honest_mf = sorted(maxflow(out, N, seeds, i) for i in range(5, 200))
    print(f"honest members' max-flow: median {honest_mf[len(honest_mf) // 2]},"
          f" range {honest_mf[0]}..{honest_mf[-1]} (scores x 1000 for PR and TR)")
    B = 100.0                                      # author's model: top-10 bonus
    need = min((k for k, r in rows if r <= 10), default=None)
    print(f"cost boundary (author's model): top-10 bonus {B:.0f}, reached under PR"
          f" with k = {need} fakes;")
    print(f"  the attack pays unless each fake identity costs more than B / k ="
          f" {B / need:.1f}")
#    k   PR score  PR rank   TR score  TR rank   MF (max-flow from seeds)
#    0     3.065      161      3.012      110     2
#    1     4.307      107      3.518       96     2
#    2     5.662       66      3.962       83     2
#    5    10.191        9      5.019       65     2
#   10    18.568        1      6.199       45     2
#   20    36.081        1      7.526       31     2
#   50    84.253        1      9.030       23     2
# honest members' max-flow: median 4, range 0..11 (scores x 1000 for PR and TR)
# cost boundary (author's model): top-10 bonus 100, reached under PR with k = 5 fakes;
#   the attack pays unless each fake identity costs more than B / k = 20.0

没有假身份时,攻击者的 PageRank 排在 201 人中的第 161 位。1 个假身份使他升到第 107 位,2 个到第 66 位,5 个就进入前十(第 9 位),10 个以上排第 1:假身份把他推荐出去的分数原样送回,PageRank 不区分谁是真的,这正是程与弗里德曼所说的对称函数的弱点。个性化 PageRank 也在上升(第 110 位升到 50 个假身份时的第 23 位),但慢得多,因为分数只能从种子经两条攻击边流进来,假身份只能让这些分数在他的小圈子里多转几圈。最大流始终是 2,恰好等于攻击边的数目:任何从种子到攻击者的路径都必须先经过攻击者本人才能到达他的假身份,所以假身份不增加流量。诚实成员的最大流中位数为 4,所以攻击者也没有因为假身份而显得更可信。

成本边界(作者的模型)。 设攻击者用 kk 个假身份可得的收益为 G(k)G(k),每个假身份的成本为 cc(注册、验证、押金被没收的期望损失、维护的时间)。攻击有利可图当且仅当 max⁡k[G(k)−ck]>G(0)\max_k\,[G(k) – c\,k] > G(0);要使任何规模的攻击都无利可图,需要

c≥supk≥1G(k)−G(0)k.c \;\ge\; \sup_{k \ge 1} \frac{G(k) – G(0)}{k}.

程序中,若排名前十可得奖励 100、PageRank 下 5 个假身份即可进入前十,则每个假身份的成本必须超过 20 才能阻止攻击。对于最大流这类防女巫函数,G(k)=G(0)G(k) = G(0),任何正的身份成本都足以阻止「制造假身份」这一种攻击,但它挡不住「骗取更多攻击边」(例如拉拢诚实成员推荐)。提高身份成本也不是免费的:弗里德曼与雷斯尼克指出,它同样落在诚实的新成员身上 (Friedman and Resnick 2001)。实名验证、押金、只从可信种子传播信任、按社交图限制身份数量,各有不同的成本与隐私代价。

左:攻击者在 201 名真实成员中的排名(对数坐标,越上越好)随假身份数量的变化:PageRank(实线)、种子个性化 PageRank(虚线),以及从种子出发的最大流(点线,右轴)。右:作者的积分模型中,五个情景下一单位奖励的积分价格。

假说 · 尚未验证 · 以流为基础的社群声誉

作者的假说(尚未确立): 在独立的线上社群中,把声誉分从「被点赞总数」改为「从一组可信成员出发的推荐流」,会使经人工审核确认的小号互赞行为显著减少,而诚实新成员在头 30 天内获得首个声誉分的比例下降不超过 10 个百分点。什么样的结果能证明它是错的: 事先登记的整群随机试验中,小号互赞的比例之差的置信区间包含零,或新成员获得声誉分的比例下降超过 10 个百分点。本书不使用作者社群的数据。

本章小结

积分是一种社群货币:按作者的模型,发放随人数增长而奖励不增长时,积分按人数增长的速度贬值(示例中月通胀 4.00%),过期只降低水平、不改变增长率,固定总量会稀释每人所得,刷分使通胀加速。徽章能改变行为:观测研究显示用户在接近阈值时加速,游戏化对学习的元分析效应为小(认知结局接近中等)、严格研究中只有认知结局稳定;简化模型显示最好的阈值取决于徽章价值与留存。对称的声誉函数不能防女巫:5 个假身份即可把攻击者的 PageRank 从第 161 位推到第 9 位,而从可信种子出发的最大流不受假身份影响。「成本边界」是作者的模型:每个假身份的成本要高于假身份带来的平均收益,才能阻止攻击。

参考文献

Anderson, Ashton, Daniel Huttenlocher, Jon Kleinberg, and Jure Leskovec. 2013. “Steering User Behavior with Badges.” Proceedings of the 22nd International Conference on World Wide Web (WWW ’13), 95–105.
Cheng, Alice, and Eric Friedman. 2005. “Sybilproof Reputation Mechanisms.” Proceedings of the 2005 ACM SIGCOMM Workshop on Economics of Peer-to-Peer Systems (P2PECON ’05), 128–32. https://doi.org/10.1145/1080192.1080202.
Douceur, John R. 2002. “The Sybil Attack.” Peer-to-Peer Systems: First International Workshop, IPTPS 2002, Lecture notes in computer science, vol. 2429: 251–60.
Fisher, Irving. 1911. The Purchasing Power of Money. Macmillan.
Friedman, Eric J., and Paul Resnick. 2001. “The Social Cost of Cheap Pseudonyms.” Journal of Economics and Management Strategy 10 (2): 173–99.
Gyöngyi, Zoltán, Hector Garcia-Molina, and Jan Pedersen. 2004. “Combating Web Spam with TrustRank.” Proceedings of the 30th International Conference on Very Large Data Bases (VLDB 2004), 576–87.
Hamari, Juho, Jonna Koivisto, and Harri Sarsa. 2014. “Does Gamification Work? A Literature Review of Empirical Studies on Gamification.” Proceedings of the 47th Hawaii International Conference on System Sciences (HICSS), 3025–34.
Lehdonvirta, Vili, and Edward Castronova. 2014. Virtual Economies: Design and Analysis. MIT Press.
Page, Lawrence, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank Citation Ranking: Bringing Order to the Web. Technical Report Nos. 1999-66. Stanford InfoLab.
Sailer, Michael, and Lisa Homner. 2020. “The Gamification of Learning: A Meta-Analysis.” Educational Psychology Review 32 (1): 77–112. https://doi.org/10.1007/s10648-019-09498-w.
Yu, Haifeng, Michael Kaminsky, Phillip B. Gibbons, and Abraham Flaxman. 2006. “SybilGuard: Defending against Sybil Attacks via Social Networks.” Proceedings of ACM SIGCOMM 2006, 267–78.

Fan, Y. (2026). Mechanism Design, Vickrey–Clarke–Groves Auctions and Incentive-Compatible Community Engineering: From Myerson's Optimal Auction and the Hurwicz Impossibility Theorem to Sybil-Resistant Reputation Games, Read Beside Classical Legalist Treatises on Rewards and Incentives [机制设计、维克里—克拉克—格罗夫斯(VCG)拍卖与数字社群激励相容工程] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23836274 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).