关于「离场风险函数:光标轨迹能预测什么,触屏上还剩什么」,《离场降单与感谢页加购》给出的定量结论是什么?

结论:截断点决定弹层打给谁。截断点取 0.3 时,弹层打中要走的人的比例,也就是召回率,是 0.517;弹出去的层里真正要走的比例,也就是精确率,是 0.495;不走的人里被误弹的比例是 0.114。截断点抬到 0.5,精确率升到 0.667,召回率跌到 0.260,误弹率降到 0.028。再抬到 0.7,精确率 0.792,召回率只剩 0.107,误弹率 0.006。

数据口径:光标轨迹确实带有意图信息,但模型输出的是概率,合成桌面数据上的 AUC 为 0.809,触屏会话失去鼠标信号后降到 0.618。 多巴胺编码的是奖赏预测误差,没有文献证明付款后存在十秒窗口;感谢页上可测的是加购接受率的半衰期,示意参数下估计为 6.50 秒,前 10 秒集中了 0.830 的接受。

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: 0.006, 0.028, 0.107, 0.114, 0.260, 0.3

sessions 4,000 test exit base rate 0.177
fitted beta-2.73 +0.94 +0.63 +0.06 +0.72
true beta-2.60 +0.90 +0.50 +0.06 +0.50
test AUC 0.809
cut pops precision recall false-pop rate
0.3370 0.495 0.517 0.114
0.5138 0.667 0.260 0.028
0.748 0.792 0.107 0.006
查看完整验算程序
# Chapter 1. p11: an exit-risk logistic model on synthetic cursor tracks.
# Model: each desktop session snapshot has four signals: cursor speed
# toward the top bar (v_up), upward acceleration (a_up), idle seconds
# (idle) and scroll stall (stall, 1 if no scroll for 5 s). The true
# chance of leaving within 2 s is logistic in these signals. We fit
# the model on half the data by Newton steps (IRLS) and score the other
# half: AUC, then precision, recall and false-pop rate at 3 cut-offs.
# All coefficients are illustrative; replace them with your own logs.
import math
import random

rng = random.Random(39811)
TRUE = [-2.6, 0.9, 0.5, 0.06, 0.5]


def sigmoid(z):
    if z < -30.0:
        return 0.0
    if z > 30.0:
        return 1.0
    return 1.0 / (1.0 + math.exp(-z))


def make(n):
    rows = []
    for _ in range(n):
        v_up = rng.gauss(0.0, 1.0)
        a_up = 0.5 * v_up + rng.gauss(0.0, 0.9)
        idle = rng.expovariate(1.0 / 6.0)
        stall = 1.0 if rng.random() < 0.35 else 0.0
        x = [1.0, v_up, a_up, idle, stall]
        z = math.fsum(b * f for b, f in zip(TRUE, x))
        y = 1 if rng.random() < sigmoid(z) else 0
        rows.append((x, y))
    return rows


def solve(a, b):
    n = len(b)
    m = [row[:] + [b[i]] for i, row in enumerate(a)]
    for c in range(n):
        piv = max(range(c, n), key=lambda r: abs(m[r][c]))
        m[c], m[piv] = m[piv], m[c]
        for r in range(n):
            if r != c:
                f = m[r][c] / m[c][c]
                m[r] = [u - f * w for u, w in zip(m[r], m[c])]
    return [m[i][n] / m[i][i] for i in range(n)]


def irls(rows, k, steps=12):
    beta = [0.0] * k
    for _ in range(steps):
        grad = [0.0] * k
        hess = [[0.0] * k for _ in range(k)]
        gterms = [[] for _ in range(k)]
        hterms = [[[] for _ in range(k)] for _ in range(k)]
        for x, y in rows:
            p = sigmoid(math.fsum(b * f for b, f in zip(beta, x)))
            w = p * (1.0 - p)
            for i in range(k):
                gterms[i].append((y - p) * x[i])
                for j in range(k):
                    hterms[i][j].append(w * x[i] * x[j])
        for i in range(k):
            grad[i] = math.fsum(gterms[i])
            for j in range(k):
                hess[i][j] = math.fsum(hterms[i][j])
        step = solve(hess, grad)
        beta = [b + s for b, s in zip(beta, step)]
    return beta


def auc(scores, labels):
    pairs = sorted(zip(scores, labels))
    rank_sum, n1 = 0, 0
    for r, (_, y) in enumerate(pairs, 1):
        if y == 1:
            rank_sum += r
            n1 += 1
    n0 = len(pairs) - n1
    return (rank_sum - n1 * (n1 + 1) / 2.0) / (n1 * n0)


data = make(4000)
train, test = data[:2000], data[2000:]
beta = irls(train, 5)
scores = [sigmoid(math.fsum(b * f for b, f in zip(beta, x)))
          for x, _ in test]
labels = [y for _, y in test]
base = math.fsum(labels) / len(labels)
print("sessions {:,}  test exit base rate {:.3f}".format(len(data), base))
print("fitted beta: " + " ".join("{:+.2f}".format(b) for b in beta))
print("true   beta: " + " ".join("{:+.2f}".format(b) for b in TRUE))
print("test AUC {:.3f}".format(auc(scores, labels)))
print("cut   pops  precision  recall  false-pop rate")
for cut in (0.3, 0.5, 0.7):
    tp = fp = fn = tn = 0
    for s, y in zip(scores, labels):
        if s >= cut:
            tp, fp = tp + y, fp + (1 - y)
        else:
            fn, tn = fn + y, tn + (1 - y)
    print("{:.1f}  {:5d}  {:9.3f}  {:6.3f}  {:14.3f}".format(
        cut, tp + fp, tp / max(1, tp + fp), tp / max(1, tp + fn),
        fp / max(1, fp + tn)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《离场轨迹微降单挽留、零门槛诱饵捕获与付款成功页黄金十秒连带母典》第 1 章「离场风险函数:光标轨迹能预测什么,触屏上还剩什么」(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-398 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026398,
  author    = {Fan, Yafu},
  title     = {Exit-Trajectory Micro-Downsells, Zero-Friction Lead Magnet Capture and Thank-You-Page Ten-Second Monetisation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论