File size: 11,620 Bytes
1b86964 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 | """Pure joint-response coupling and recorded semantic corrections.
Shared empirical ranks are a model prior, not recovered human correlations.
All indices are zero based; answer option IDs are one based. Inputs are never
modified, and this module performs no file or network access.
"""
from __future__ import annotations
import copy
import json
import re
import numpy as np
from scipy.optimize import linear_sum_assignment
from scipy.spatial.distance import cdist
from scipy.stats import rankdata
def components(inp):
"""Keep items used by one canonical statistic in an indivisible group."""
n = len(inp["prompts"][0]["questions"])
parent = list(range(n))
used = set()
def root(i):
while parent[i] != i:
parent[i] = parent[parent[i]]
i = parent[i]
return i
def join(items):
items = sorted(set(items))
assert all(0 <= j < n for j in items)
used.update(items)
for j in items[1:]:
parent[root(j)] = root(items[0])
for effect in inp["effects"]:
metadata = {"identifiability": effect["identifiability"],
"statistic_spec": effect["statistic_spec"]}
join([int(s) - 1 for s in re.findall(r"item_(\d+)", json.dumps(metadata))])
for block in inp["blocks"]:
join([j for j, _ in block])
join(inp["categorical"])
groups = {}
for j in sorted(used):
groups.setdefault(root(j), []).append(j)
free = sorted(set(range(n)) - used)
return sorted(groups.values(), key=lambda x: x[0]), free
def rank_matrix(a):
return rankdata(a, axis=0, method="average") / (len(a) + 1.)
def joint_diagnostic(a, reference, groups, free):
blocks = groups + ([free] if free else [])
owner = {j: k for k, block in enumerate(blocks) for j in block}
with np.errstate(divide="ignore", invalid="ignore"):
ca = np.atleast_2d(np.corrcoef(rank_matrix(a), rowvar=False))
cr = np.atleast_2d(np.corrcoef(rank_matrix(reference), rowvar=False))
diffs = []
inversions = 0
for i in range(a.shape[1]):
for j in range(i):
if owner[i] == owner[j]:
continue
if np.isfinite(ca[i, j]) and np.isfinite(cr[i, j]):
diffs.append(abs(float(ca[i, j] - cr[i, j])))
inversions += ca[i, j] * cr[i, j] < 0 and min(abs(ca[i, j]), abs(cr[i, j])) > .1
return {"between_component_rank_correlation_MAE": float(np.mean(diffs)) if diffs else None,
"correlation_pairs": len(diffs),
"opposite_sign_pairs_above_0_1": int(inversions),
"unique_joint_vectors": len(np.unique(a, axis=0)),
"constant_items": int((np.std(a, axis=0) == 0).sum())}
def couple(prompt, a, initial, groups, free, rng):
"""Transport whole fitted components to a shared initial-model rank prior."""
n, _ = a.shape
assert initial.shape == a.shape
proposals = []
for _ in range(4):
out = initial.copy()
permutations = []
cost = 0.
for block in groups:
if len(groups) == 1 and not free:
ix = np.arange(n)
else:
target = rank_matrix(initial[:, block])
donors = rank_matrix(a[:, block])
distances = cdist(target, donors, "sqeuclidean") / len(block)
_, ix = linear_sum_assignment(distances + rng.uniform(0, 1e-10, distances.shape))
cost += float(distances[np.arange(n), ix].mean())
out[:, block] = a[ix][:, block]
permutations.append(ix)
assert sorted(map(tuple, out[:, block])) == sorted(map(tuple, a[:, block]))
assert not free or np.array_equal(out[:, free], initial[:, free])
diag = joint_diagnostic(out, initial, groups, free)
score = diag["between_component_rank_correlation_MAE"]
proposals.append((float(score) if score is not None else 0., cost, out, permutations, diag))
_, _, out, permutations, after = min(proposals, key=lambda x: x[:2])
before = joint_diagnostic(a, initial, groups, free)
retained = False
if (not free and before["between_component_rank_correlation_MAE"] is not None
and after["between_component_rank_correlation_MAE"]
> before["between_component_rank_correlation_MAE"] + 1e-10):
out = a.copy()
permutations = [np.arange(n) for _ in groups]
retained = True
result = {key: copy.deepcopy(prompt[key]) for key in
("study_id", "condition_key", "condition_id", "original_n", "questions", "messages")}
result["respondents"] = out.tolist()
result["protected_components"] = copy.deepcopy(groups)
result["free_items"] = list(free)
result["retained_coupling"] = retained
result["respondent_provenance"] = []
for i in range(n):
parts = [{"items": list(block), "source": "fitted",
"condition_key": prompt["condition_key"], "respondent_index": int(ix[i])}
for block, ix in zip(groups, permutations)]
if free:
parts.append({"items": list(free), "source": "initial",
"condition_key": prompt["condition_key"], "respondent_index": i})
result["respondent_provenance"].append({"components": parts})
return result
def balance_audience(populations):
"""Preserve pooled response vectors and n; audience balance is a prior."""
by = {q["condition_id"]: q for q in populations}
rng = np.random.default_rng(20260908)
for left, right in [("A", "B"), ("C", "D"), ("E", "F")]:
p, q = by[left], by[right]
n1, n2 = p["original_n"], q["original_n"]
x, y = np.array(p["respondents"]), np.array(q["respondents"])
x[:, 0] = 4 - x[:, 0]
y[:, 0] = y[:, 0] - 4
pool = np.concatenate([x, y])
provenance = p["respondent_provenance"] + q["respondent_provenance"]
features = np.column_stack([pool, pool ** 2]).astype(float)
features /= np.maximum(features.std(0), 1.)
target = features.sum(0) * n1 / (n1 + n2)
best = np.arange(n1 + n2)
best_cost = float("inf")
for _ in range(1000):
ix = rng.permutation(n1 + n2)
err = features[ix[:n1]].sum(0) - target
cost = float(err @ err)
if cost < best_cost:
best, best_cost = ix, cost
err = features[best[:n1]].sum(0) - target
for _ in range(20000):
i = int(rng.integers(n1))
j = int(rng.integers(n1, n1 + n2))
trial = err - features[best[i]] + features[best[j]]
cost = float(trial @ trial)
if cost < best_cost - 1e-12:
best[i], best[j] = best[j], best[i]
err = trial
best_cost = cost
a = pool[best[:n1]].copy()
b = pool[best[n1:]].copy()
assert sorted(map(tuple, pool)) == sorted(map(tuple, np.concatenate([a, b])))
a[:, 0] = 4 - a[:, 0]
b[:, 0] = b[:, 0] + 4
for dest, values, indices in [(p, a, best[:n1]), (q, b, best[n1:])]:
dest["respondents"] = values.tolist()
new_provenance = []
for ix in indices:
pr = copy.deepcopy(provenance[int(ix)])
source_condition = left if ix < n1 else right
pr["audience_reassignment"] = {
"from_condition": source_condition, "to_condition": dest["condition_id"],
"recoded_item": 0,
"option_mapping": "8-option_id" if source_condition != dest["condition_id"] else "identity"}
new_provenance.append(pr)
dest["respondent_provenance"] = new_provenance
def repair_prompts(populations, policy):
"""Apply recorded wording corrections supplied with the study inputs."""
for row in populations:
for edit in policy["wording_corrections"]:
selectors = {key: edit[key] for key in ("condition_key", "study_id") if key in edit}
if not selectors:
raise ValueError("A wording correction must identify its study or condition")
if any(row[key] != value for key, value in selectors.items()):
continue
text = row["messages"][1]["content"]
if edit["only_if_present"] and edit["match"] not in text:
continue
assert text.count(edit["match"]) == 1
assert all(fragment in text for fragment in edit["required"])
row["messages"][1]["content"] = text.replace(edit["before"], edit["after"])
def semantic_rules(populations, free_items, policy):
"""Correct only explicitly specified free-item facts, not human frequencies."""
for q in populations:
key = q["condition_key"]
for edit in policy["prompt_edits"]:
if edit["condition_key"] != key:
continue
text = q["messages"][1]["content"]
assert text.count(edit["before"]) == 1, (key, "stimulus source changed")
q["messages"][1]["content"] = text.replace(edit["before"], edit["after"])
for rule in policy["facts"]:
if rule["condition_key"] != key:
continue
j, value = rule["item_index"], rule["expected_option_id"]
assert j in free_items, (key, j, "cannot alter canonical dependency")
assert rule["question"] in q["messages"][1]["content"]
assert q["questions"][j]["options"][value - 1]["text"] == rule["expected_option_text"]
for answer, lineage in zip(q["respondents"], q["respondent_provenance"]):
if answer[j] == value:
continue
lineage.setdefault("semantic_overrides", []).append({
"rule_id": rule["rule_id"], "item_index": j,
"before_option_id": answer[j], "after_option_id": value})
answer[j] = value
q["semantic_rules"] = [r["rule_id"] for r in policy["facts"] if r["condition_key"] == key]
def finalize_study(inp, respondents, recipe, policy):
"""Finalize fitted responses without reading saved final populations.
Preserve prompt order and the shared per-study RNG stream. The caller
supplies integer option matrices after construct/moment transformations.
"""
groups, free = components(inp)
rng = np.random.default_rng(int(recipe["transport_seed"]))
expected_keys = {q["condition_key"] for q in inp["prompts"]}
assert len(expected_keys) == len(inp["prompts"])
assert set(respondents) == expected_keys
populations = []
for prompt in inp["prompts"]:
assert prompt["study_id"] == inp["study"]
key = prompt["condition_key"]
a = np.array(respondents[key], dtype=int)
assert np.array_equal(a, np.asarray(respondents[key]))
assert a.shape == (prompt["original_n"], len(prompt["questions"]))
initial = np.array(inp["sources"]["Chiron"][key], dtype=int)
populations.append(couple(prompt, a, initial, groups, free, rng))
if inp["study"] == "JEP_1_6b":
balance_audience(populations)
repair_prompts(populations, policy)
semantic_rules(populations, free, policy)
for q in populations:
a = np.array(q["respondents"])
assert a.shape == (q["original_n"], len(q["questions"]))
assert all(1 <= int(value) <= len(q["questions"][j]["options"])
for row in a for j, value in enumerate(row))
assert len(q["respondent_provenance"]) == q["original_n"]
return populations
|