What does Hidden Markov Models for Customer Decision Stages conclude, in numbers, about “Identifiability and Label Switching: What Clickstreams Cannot Say”?
Answer: Chapter 5 treats label switching, the number of stages and weak identifiability: naming stages by average position within a visit gives an accuracy of 0.4477, close to 0.4623 with true labels, and with blurred pages clicks keep only 0.022 of the information.
Scope of the numbers: Marked boxes separate published evidence, runnable Python calculations (18 programs), classical sources and the author's hypotheses.
Data table
The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.4477, 0.4623
| part 1 | 24 relabellings of the true model, p21 visit |
| log P(clicks) min -12.588330510522 max -12.588330510522 | |
| part 2: runs of p42; naming rules against the oracle | |
| run oracle E P acc: orc E P ready: E P | |
| 1 | 3012 3021 3021 0.4303 0.4277 0.4277 0.7181 0.7181 |
| 2 | 0213 0123 0213 0.5039 0.4910 0.5039 0.7406 0.7716 |
| 3 | 0123 1023 0213 0.4845 0.3026 0.4465 0.7813 0.7413 |
| 4 | 2301 2310 2301 0.4619 0.4226 0.4619 0.7813 0.7813 |
| 5 | 3210 2130 3120 0.3877 0.3600 0.3787 0.6458 0.6458 |
| 6 | 0321 0312 0312 0.4129 0.4032 0.4032 0.7413 0.7413 |
| 7 | 3201 2301 1302 0.4329 0.4303 0.3910 0.7129 0.7097 |
| 8 | 3012 3102 3021 0.5181 0.3065 0.4974 0.7703 0.7471 |
| 9 | 1230 1230 2130 0.5123 0.5123 0.4903 0.7884 0.7587 |
| 10 | 3201 3012 3102 0.4787 0.3219 0.4768 0.7703 0.7703 |
| mean 4-stage acc: oracle 0.4623, rule E 0.3978, rule P 0.4477 | |
| rule E matches the oracle in 1 of 10 runs, rule P in 2 |
Show the full program
# Chapter 5. p51: label switching, and two ways to name fitted stages.
# Model: the four-stage HMM of p11 (illustrative). Part 1: permuting the
# stage labels of the true model (all 24 orders) leaves the likelihood
# of the p21 visit unchanged. Part 2: the same 150 visits and the same
# 10 random starts as p42 (same seed, 40 Baum-Welch iterations). The
# fitted states carry arbitrary numbers. Two naming rules that need no
# true labels are compared with the oracle relabelling of p42:
# rule E orders states by their share of SHIP and CART pages; rule P
# orders states by mean posterior click position within a visit.
import itertools
import math
import random
TRUE_PI = [0.70, 0.20, 0.08, 0.02]
TRUE_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]]
TRUE_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, M = 4, 6
VISIT = [0, 1, 2, 3, 1, 3, 4, 5]
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 rand_row(rng, n):
w = [rng.random() + 0.2 for _ in range(n)]
z = math.fsum(w)
return [x / z for x in w]
def fb(pi, a, b, obs):
T = len(obs)
al, cs = [], []
x = [pi[k] * b[k][obs[0]] for k in range(K)]
for t in range(T):
if t:
x = [math.fsum(al[-1][i] * a[i][j] for i in range(K))
* b[j][obs[t]] for j in range(K)]
c = math.fsum(x)
cs.append(c)
al.append([v / c for v in x])
be = [[1.0] * K]
for t in range(T - 2, -1, -1):
nb = be[0]
be.insert(0, [math.fsum(a[i][j] * b[j][obs[t + 1]] * nb[j]
for j in range(K)) / cs[t + 1]
for i in range(K)])
gam = [[al[t][k] * be[t][k] for k in range(K)] for t in range(T)]
xi = [[[al[t][i] * a[i][j] * b[j][obs[t + 1]] * be[t + 1][j]
/ cs[t + 1] for j in range(K)] for i in range(K)]
for t in range(T - 1)]
return math.fsum(math.log(c) for c in cs), gam, xi
def em_step(pi, a, b, data):
ll, g1 = [], [[] for _ in range(K)]
na = [[[] for _ in range(K)] for _ in range(K)]
da = [[] for _ in range(K)]
nb = [[[] for _ in range(M)] for _ in range(K)]
db = [[] for _ in range(K)]
for obs in data:
lv, gam, xi = fb(pi, a, b, obs)
ll.append(lv)
for k in range(K):
g1[k].append(gam[0][k])
for t, o in enumerate(obs):
for k in range(K):
nb[k][o].append(gam[t][k])
db[k].append(gam[t][k])
if t < len(obs) - 1:
da[k].append(gam[t][k])
for j in range(K):
na[k][j].append(xi[t][k][j])
n = len(data)
pi = [math.fsum(g1[k]) / n for k in range(K)]
a = [[math.fsum(na[i][j]) / math.fsum(da[i]) for j in range(K)]
for i in range(K)]
b = [[math.fsum(nb[k][m]) / math.fsum(db[k]) for m in range(M)]
for k in range(K)]
return pi, a, b, math.fsum(ll)
def viterbi(pi, a, b, obs):
d = [math.log(pi[k] + 1e-300) + math.log(b[k][obs[0]] + 1e-300)
for k in range(K)]
back = []
for o in obs[1:]:
ptr = [max(range(K), key=lambda i: d[i] + math.log(a[i][j]
+ 1e-300))
for j in range(K)]
d = [d[ptr[j]] + math.log(a[ptr[j]][j] + 1e-300)
+ math.log(b[j][o] + 1e-300) for j in range(K)]
back.append(ptr)
path = [max(range(K), key=lambda k: d[k])]
for ptr in reversed(back):
path.append(ptr[path[-1]])
return path[::-1]
def rank_to_perm(score):
# 分数从小到大依次命名为 BRW、RES、CMP、DEC
order = sorted(range(K), key=lambda k: (score[k], k))
perm = [0] * K
for name, k in enumerate(order):
perm[k] = name
return tuple(perm)
lls = []
for perm in itertools.permutations(range(K)):
inv = [perm.index(k) for k in range(K)]
pi = [TRUE_PI[inv[k]] for k in range(K)]
a = [[TRUE_A[inv[i]][inv[j]] for j in range(K)] for i in range(K)]
b = [TRUE_B[inv[k]] for k in range(K)]
lls.append(fb(pi, a, b, VISIT)[0])
print("part 1: 24 relabellings of the true model, p21 visit")
print(" log P(clicks) min %.12f max %.12f" % (min(lls), max(lls)))
rng = random.Random(20261113)
data, truth = [], []
for _ in range(150):
s = [draw(rng, TRUE_PI)]
for _ in range(rng.randint(6, 14) - 1):
s.append(draw(rng, TRUE_A[s[-1]]))
truth.append(s)
data.append([draw(rng, TRUE_B[k]) for k in s])
print("part 2: runs of p42; naming rules against the oracle")
print("run oracle E P acc: orc E P ready: E P")
ok_e = ok_p = 0
acc_e, acc_p, acc_o = [], [], []
def score(paths, perm):
pairs = [(perm[x], y) for p, s in zip(paths, truth)
for x, y in zip(p, s)]
n = len(pairs)
hit = len([1 for x, y in pairs if x == y]) / n
rdy = len([1 for x, y in pairs if (x >= 2) == (y >= 2)]) / n
return hit, rdy
for r in range(10):
pi = rand_row(rng, K)
a = [rand_row(rng, K) for _ in range(K)]
b = [rand_row(rng, M) for _ in range(K)]
for _ in range(40):
pi, a, b, _ = em_step(pi, a, b, data)
paths = [viterbi(pi, a, b, o) for o in data]
oracle, best_hit = None, -1
for perm in itertools.permutations(range(K)):
hit = len([1 for p, s in zip(paths, truth)
for x, y in zip(p, s) if perm[x] == y])
if hit > best_hit:
oracle, best_hit = perm, hit
pe = rank_to_perm([b[k][4] + b[k][5] for k in range(K)])
pos = [[] for _ in range(K)]
wts = [[] for _ in range(K)]
for obs in data:
gam = fb(pi, a, b, obs)[1]
for t in range(len(obs)):
for k in range(K):
pos[k].append(gam[t][k] * t / (len(obs) - 1))
wts[k].append(gam[t][k])
pp = rank_to_perm([math.fsum(pos[k]) / math.fsum(wts[k])
for k in range(K)])
ok_e += 1 if pe == oracle else 0 # step
ok_p += 1 if pp == oracle else 0 # step
so, se, sp = score(paths, oracle), score(paths, pe), score(paths, pp)
acc_o.append(so[0])
acc_e.append(se[0])
acc_p.append(sp[0])
print("%2d %s %s %s %.4f %.4f %.4f %.4f %.4f"
% (r + 1, "".join(map(str, oracle)), "".join(map(str, pe)),
"".join(map(str, pp)), so[0], se[0], sp[0], se[1], sp[1]))
print("mean 4-stage acc: oracle %.4f, rule E %.4f, rule P %.4f"
% (math.fsum(acc_o) / 10, math.fsum(acc_e) / 10,
math.fsum(acc_p) / 10))
print("rule E matches the oracle in %d of 10 runs, rule P in %d"
% (ok_e, ok_p))Single source and citation
Fan, Y. (2026). Hidden Markov Models and Viterbi Decoding: Inferring a Customer's Hidden Decision Stage from Click Sequences and Timing the Offer (Version 1.0.0), Chapter 5: Identifiability and Label Switching: What Clickstreams Cannot Say. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 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}
}Other chapter answers in this book
- Chapter 1 · The Last-Click Illusion: Whole Sequences against a Single Page
- Chapter 2 · The Viterbi Trellis: Dynamic Programming in Logarithms, Backtracking and Underflow
- Chapter 3 · Forward and Backward: Posterior Marginals, the Best Path and "Wait Two More Clicks"
- Chapter 4 · Baum–Welch: Estimating Parameters from Unlabelled Clickstreams
- Chapter 6 · Timing the Offer: the Inquiry as Evidence, a Cost Threshold and Calibration
- Chapter 7 · Classical Sources East and West: Markov's Letter Chain in Onegin and the Chuai Chapter of the Guiguzi
- Chapter 8 · Closing the Loop: a Weekly Stage Review, the Customer Stage Decoder and Pre-Registered Tests