关于「维特比格子:对数域动态规划、回溯与下溢」,《隐马尔可夫与客户决策阶段》给出的定量结论是什么?
结论:第二章在格子图上展开维特比,说明直接相乘约在第 368 次点击下溢,最优路径只占全部概率的 0.0357。
数据口径:书中阶段、页面类型与全部概率均为示意;全部计算由书中程序生成;本书是非个人化经营建议,只使用站内匿名化与客户同意的数据,不跨站追踪,不推断敏感属性,阶段解码不用于施压、差别定价或主动联系客户,紧迫感只来自真实库存与真实截止。 第一章在示意参数下量出:前向滤波的四分类准确率比最后一次点击高 0.0410,就绪判别从 0.6990 提高到 0.7630;增益随阶段黏性与页面区分力在 0.0307 到 0.1187 之间变化,维特比最后一格在页面模糊时反而低 0.1007。
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: 368
| smallest positive double: 4.941e-324 | ||||
| T | raw log P | log-domain | scaled | raw path = log path |
| 50 | -105.41 | -105.41 | -105.41 | 1.000 |
| 200 | -410.93 | -410.93 | -410.93 | 1.000 |
| 400 | -inf(0.0) | -818.24 | -818.24 | 0.943 |
| 600 | -inf(0.0) | -1228.56 | -1228.56 | 0.632 |
| 1000 | -inf(0.0) | -2042.15 | -2042.15 | 0.408 |
| 2000 | -inf(0.0) | -4042.20 | -4042.20 | 0.255 |
| log-probability per click: -2.021 nats | ||||
| raw products die near T = 368 clicks |
核心公式
\[\delta_t(j) = \max_i \left[ \delta_{t-1}(i)\, a_{ij} \right] b_j(o_t),\]
查看完整验算程序
# Chapter 2. p22: why decoding must run in logs (or with scaling).
# Model: the four-stage HMM of p11 (illustrative). One long click stream
# is simulated with a fixed seed, as if a member's consented, anonymised
# clicks over many visits were decoded together. At several lengths T
# the program runs the Viterbi recursion three ways: raw products of
# probabilities, natural logs, and per-step rescaling (divide by the
# column maximum and keep the log of the divisor). Raw products sink
# below the smallest double and the decoded path turns to nonsense.
import math
import random
PI = [0.70, 0.20, 0.08, 0.02]
A = [[0.70, 0.22, 0.06, 0.02],
[0.08, 0.64, 0.23, 0.05],
[0.02, 0.10, 0.66, 0.22],
[0.01, 0.03, 0.16, 0.80]]
B = [[0.45, 0.30, 0.08, 0.07, 0.06, 0.04],
[0.15, 0.35, 0.25, 0.12, 0.08, 0.05],
[0.08, 0.22, 0.15, 0.30, 0.17, 0.08],
[0.05, 0.12, 0.08, 0.15, 0.30, 0.30]]
K = 4
def draw(rng, probs):
u, acc = rng.random(), 0.0
for k, p in enumerate(probs):
acc += p # step
if u < acc:
return k
return len(probs) - 1
def decode(obs, mode):
# mode: raw 原始乘积;log 对数域;scale 每步按列最大值缩放
if mode == "log":
d = [math.log(PI[k] * B[k][obs[0]]) for k in range(K)]
else:
d = [PI[k] * B[k][obs[0]] for k in range(K)]
logc = []
back = []
for o in obs[1:]:
if mode == "log":
ptr = [max(range(K), key=lambda i: d[i] + math.log(A[i][j]))
for j in range(K)]
d = [d[ptr[j]] + math.log(A[ptr[j]][j] * B[j][o])
for j in range(K)]
else:
ptr = [max(range(K), key=lambda i: d[i] * A[i][j])
for j in range(K)]
d = [d[ptr[j]] * A[ptr[j]][j] * B[j][o] for j in range(K)]
if mode == "scale":
c = max(d)
logc.append(math.log(c))
d = [x / c for x in d]
back.append(ptr)
best = max(range(K), key=lambda k: d[k])
path = [best]
for ptr in reversed(back):
path.append(ptr[path[-1]])
if mode == "log":
score = d[best]
elif mode == "scale":
score = math.log(d[best]) + math.fsum(logc)
else:
score = math.log(d[best]) if d[best] > 0 else -math.inf
return path[::-1], score
rng = random.Random(20261113)
s = [0]
for _ in range(1999):
s.append(draw(rng, A[s[-1]]))
o = [draw(rng, B[k]) for k in s]
print("smallest positive double: %.3e" % (5e-324))
print(" T raw log P log-domain scaled raw path = log path")
for t_len in (50, 200, 400, 600, 1000, 2000):
obs = o[:t_len]
p_raw, sc_raw = decode(obs, "raw")
p_log, sc_log = decode(obs, "log")
p_sc, sc_sc = decode(obs, "scale")
agree = len([1 for a, b in zip(p_raw, p_log) if a == b]) / t_len
raw_txt = "%10.2f" % sc_raw if sc_raw > -math.inf else " -inf(0.0)"
print("%5d %s %11.2f %8.2f %8.3f%s"
% (t_len, raw_txt, sc_log, sc_sc, agree,
"" if p_sc == p_log else " scaled path differs"))
per_click = sc_log / 2000
print("log-probability per click: %.3f nats" % per_click)
print("raw products die near T = %d clicks"
% int(math.log(5e-324) / per_click))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《隐马尔可夫维特比解码、浏览点击序列反推客户隐藏决策阶段与出手时机母典》第 2 章「维特比格子:对数域动态规划、回溯与下溢」(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-300 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026300,
author = {Fan, Yafu},
title = {Hidden Markov Models and Viterbi Decoding: Inferring a Customer's Hidden Decision Stage from Click Sequences and Timing the Offer},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}