关于「缺失非随机:只看到买了的,会把推断带偏多少」,《稀疏购买记录的偏好补全》给出的定量结论是什么?

结论:第五章说明只看到买了的会怎样:选择偏强时未买格子的预测偏高 0.4245,倾向加权降到 0.1671;只拟合已买格子的模型命中率仅 0.286,把没买的格子也计入的加权交替最小二乘达到 0.743。

数据口径:书中顾客数、商品数、口味维度与购买概率均为示意参数;全部计算由书中程序生成;本书属于非个人化经营建议,推荐流程只用店铺自有数据,不外传个人数据,不做跨平台追踪。 第七章考辨《周易·系辞上》「方以类聚,物以群分」与斯皮尔曼 1904 年的一般因素,指出纯分群与混合成员下两种模型各有胜负。

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: 0.1671, 0.4245

b = 2, seen 299, true unseen mean 2.6881
methodweightmean estunseen biasunseen RMSE
plainess2993.7260+0.42450.8633
ipw cap 20ess1583.1522+0.16710.2619
查看完整验算程序
# Chapter 5. p52: inverse-propensity weights against selection bias.
# Model: the store of p51 with selection strength b = 2. When the
# purchase probability pi of each cell is known (in practice it must
# be estimated), weighting each observed cell by 1/pi in ALS and in
# the mean undoes the selection on average. Weights are capped at 20.
# Effective sample size (sum w)^2 / sum w^2 shows the price paid.
# All parameters illustrative.
import math
import random


def dot(a, b):
    return math.fsum(x * y for x, y in zip(a, b))


def solve(A, b):
    # Gaussian elimination with partial pivoting for a small system
    n = len(b)
    M = [list(A[i]) + [b[i]] for i in range(n)]
    for k in range(n):
        p = max(range(k, n), key=lambda i: abs(M[i][k]))
        M[k], M[p] = M[p], M[k]
        for i in range(k + 1, n):
            f = M[i][k] / M[k][k]
            M[i] = [M[i][j] - f * M[k][j] for j in range(n + 1)]
    x = [0.0] * n
    for i in range(n - 1, -1, -1):
        s = math.fsum(M[i][j] * x[j] for j in range(i + 1, n))
        x[i] = (M[i][n] - s) / M[i][i]
    return x


def update(trip, F, r, lam, prior=None):
    # weighted ridge step for one row: minimise
    # sum w (y - f.x)^2 + lam |x - prior|^2 over x
    pr = prior if prior is not None else [0.0] * r
    A = [[math.fsum(w * F[j][a] * F[j][b] for j, _, w in trip)
          + (lam if a == b else 0.0) for b in range(r)]
         for a in range(r)]
    y = [math.fsum(w * F[j][a] * v for j, v, w in trip) + lam * pr[a]
         for a in range(r)]
    return solve(A, y)


def als(nu, ni, obs, r, lam, iters, seed, V=None, trace=None):
    # alternating least squares on (u, i, value, weight) tuples
    rng = random.Random(seed)
    if V is None:
        V = [[rng.gauss(0.0, 1.0) for _ in range(r)] for _ in range(ni)]
    bu = [[] for _ in range(nu)]
    bi = [[] for _ in range(ni)]
    for u, i, x, w in obs:
        bu[u].append((i, x, w))
        bi[i].append((u, x, w))
    U = None
    for t in range(iters):
        U = [update(bu[u], V, r, lam) for u in range(nu)]
        V = [update(bi[i], U, r, lam) for i in range(ni)]
        if trace is not None:
            trace.append((t + 1, U, V))
    return U, V


def factors(n, r, rng):
    return [[rng.gauss(0.0, 1.0) for _ in range(r)] for _ in range(n)]


def product(A, B):
    return [[dot(a, b) for b in B] for a in A]


def mask(nu, ni, p, rng):
    return [(u, i) for u in range(nu) for i in range(ni)
            if rng.random() < p]

def logistic(x):
    return 1.0 / (1.0 + math.exp(-x))


rng = random.Random(20261113)
nu, ni, r, b = 40, 30, 2, 2.0
F = product([[0.7 * x for x in row] for row in factors(nu, r, rng)],
            factors(ni, r, rng))
T = [[3.0 + x for x in row] for row in F]
cells = [(u, i) for u in range(nu) for i in range(ni)]
mu = math.fsum(T[u][i] for u, i in cells) / len(cells)
sd = math.sqrt(math.fsum((T[u][i] - mu) ** 2 for u, i in cells)
               / len(cells))
draws = [rng.random() for _ in cells]
z = [(T[u][i] - mu) / sd for u, i in cells]
lo, hi = -10.0, 10.0
for _ in range(80):
    a = (lo + hi) / 2
    share = math.fsum(logistic(a + b * x) for x in z) / len(z)
    lo, hi = (a, hi) if share < 0.25 else (lo, a)
pi = {c: logistic(a + b * x) for c, x in zip(cells, z)}
seen = [c for c, d in zip(cells, draws) if d < pi[c]]
ss = set(seen)
rest = [c for c in cells if c not in ss]
true = math.fsum(T[u][i] for u, i in rest) / len(rest)
print("b = 2, seen {}, true unseen mean {:.4f}".format(len(seen), true))
print(" method        weight    mean est   unseen bias   unseen RMSE")
for name, cap in (("plain", None), ("ipw cap 20", 20.0)):
    wt = {c: 1.0 if cap is None else min(cap, 1.0 / pi[c]) for c in seen}
    W = math.fsum(wt.values())
    m0 = math.fsum(wt[c] * T[c[0]][c[1]] for c in seen) / W
    obs = [(u, i, T[u][i] - m0, wt[(u, i)]) for u, i in seen]
    U, V = als(nu, ni, obs, r, 0.1, 60, 4)
    pr = [m0 + dot(U[u], V[i]) for u, i in rest]
    bias = math.fsum(pr) / len(pr) - true
    err = math.sqrt(math.fsum((x - T[u][i]) ** 2
                              for x, (u, i) in zip(pr, rest)) / len(pr))
    ess = W * W / math.fsum(x * x for x in wt.values())
    print(" {:11s}  ess {:4.0f}   {:.4f}     {:+.4f}       {:.4f}".format(
        name, ess, m0, bias, err))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《低秩矩阵补全与交替最小二乘、稀疏购买记录推断未表态偏好母典》第 5 章「缺失非随机:只看到买了的,会把推断带偏多少」(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-328 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026328,
  author    = {Fan, Yafu},
  title     = {Low-Rank Matrix Completion and Alternating Least Squares: Inferring Unstated Preferences from Sparse Purchase Records},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论