Spaces:
Running
Running
File size: 19,236 Bytes
e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 c8c01fa 0f5b08b c8c01fa 0f5b08b c8c01fa e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 7cbf97d e5b7184 7cbf97d 0f5b08b 7cbf97d 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 7cbf97d 0f5b08b e5b7184 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b c8c01fa a3849be 0f5b08b a3849be 0f5b08b a3849be 0f5b08b 7cbf97d 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b a3849be 0f5b08b a3849be 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b c8c01fa 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b 7cbf97d e5b7184 a3849be e5b7184 7cbf97d e5b7184 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b e5b7184 0f5b08b c8c01fa 7cbf97d 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b 7cbf97d 0f5b08b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 | """
tools.py
--------
Topic-modelling pipeline: SPECTER-2 → UMAP → HDBSCAN
with multi-objective Bayesian optimisation over UMAP + HDBSCAN
parameters (§3.1–§3.6 of the methodology guide).
No BERTopic wrapper — bare UMAP + HDBSCAN on SPECTER-2 embeddings.
"""
import re
import logging
import pandas as pd
import numpy as np
from typing import Optional
from collections import Counter, defaultdict
from sentence_transformers import SentenceTransformer
from umap import UMAP
from hdbscan import HDBSCAN
from sklearn.metrics import adjusted_rand_score
from sklearn.metrics.pairwise import cosine_similarity
import optuna
# ---------------------------------------------------------------------------
# Logging
# ---------------------------------------------------------------------------
logging.basicConfig(level=logging.INFO, format="%(levelname)s | %(message)s")
logger = logging.getLogger(__name__)
optuna.logging.set_verbosity(optuna.logging.WARNING)
# ---------------------------------------------------------------------------
# Data Loading (unchanged)
# ---------------------------------------------------------------------------
def load_csv(filepath: str) -> pd.DataFrame:
df = pd.read_csv(filepath)
df.columns = df.columns.str.lower()
required = {"title", "abstract"}
missing = required - set(df.columns)
if missing:
raise ValueError(f"CSV missing column(s): {missing}")
logger.info("Loaded %d rows from '%s'.", len(df), filepath)
return df
# ---------------------------------------------------------------------------
# §3.1 — Input unit: title + abstract concatenation
# ---------------------------------------------------------------------------
def prepare_documents(df: pd.DataFrame) -> list[str]:
"""One string per paper: title + abstract (§3.1 input unit)."""
docs = (df["title"].fillna("") + ". " + df["abstract"].fillna("")).tolist()
logger.info("Prepared %d title+abstract documents.", len(docs))
return docs
# ---------------------------------------------------------------------------
# §3.1 — Embed with SPECTER-2 (cached model for speed)
# ---------------------------------------------------------------------------
_MODEL_CACHE = {}
def embed_documents(
docs: list[str],
model_name: str = "allenai/specter2_base",
) -> np.ndarray:
"""Embed with SPECTER-2. Deterministic — no tuning (§3.3)."""
if model_name not in _MODEL_CACHE:
logger.info("Loading %s (first time, will be cached)…", model_name)
_MODEL_CACHE[model_name] = SentenceTransformer(model_name)
model = _MODEL_CACHE[model_name]
embeddings = model.encode(docs, show_progress_bar=True, batch_size=64)
logger.info("Embedded %d docs → %s", len(docs), embeddings.shape)
return embeddings
# ---------------------------------------------------------------------------
# §3.2 — Cluster discipline checks
# ---------------------------------------------------------------------------
def check_discipline(labels: np.ndarray, n_docs: int) -> dict:
"""Two hard constraints: max-mass ≤ 25 %, min-size ≥ 5."""
counts = Counter(int(l) for l in labels)
unique = [l for l in counts if l != -1]
if not unique:
return dict(max_mass_pct=0, max_mass_ok=False,
min_size=0, min_size_ok=False,
n_clusters=0, n_noise=counts.get(-1, 0))
max_mass_pct = max(counts[l] / n_docs for l in unique)
min_size = min(counts[l] for l in unique)
return dict(
max_mass_pct=round(max_mass_pct, 4),
max_mass_ok=max_mass_pct <= 0.25,
min_size=int(min_size),
min_size_ok=min_size >= 5,
n_clusters=len(unique),
n_noise=counts.get(-1, 0),
cluster_sizes={l: counts[l] for l in sorted(unique)},
)
# ---------------------------------------------------------------------------
# §3.4 — Quality metrics
# ---------------------------------------------------------------------------
def compute_persistence(clusterer: HDBSCAN) -> float:
"""Average cluster persistence from the condensed tree."""
try:
p = getattr(clusterer, "cluster_persistence_", None)
if p is not None and len(p) > 0:
return float(np.mean(p))
except Exception:
pass
return 0.0
def per_cluster_persistence(clusterer: HDBSCAN, labels: np.ndarray) -> dict:
"""Map each cluster ID to its persistence score (§8)."""
try:
p = getattr(clusterer, "cluster_persistence_", None)
if p is None or len(p) == 0:
return {}
unique = sorted(set(int(l) for l in labels if l != -1))
return {cid: float(p[i]) if i < len(p) else 0.0
for i, cid in enumerate(unique)}
except Exception:
return {}
def compute_dbcv(reduced: np.ndarray, labels: np.ndarray) -> float:
"""Density-Based Cluster Validity index."""
try:
from hdbscan.validity import validity_index
ul = set(labels); ul.discard(-1)
if len(ul) < 2:
return -1.0
return float(validity_index(reduced.astype(np.float64), labels))
except Exception as e:
logger.warning("DBCV failed: %s", e)
return -1.0
def compute_stability(embeddings: np.ndarray, params: dict,
n_seeds: int = 3) -> float:
"""Cluster-recurrence stability via pairwise ARI across seeds (§3.4).
Uses 3 seeds by default for speed (spec allows 3–5)."""
all_labels = []
for s in range(n_seeds):
u = UMAP(n_neighbors=params["n_neighbors"],
n_components=params["n_components"],
min_dist=0.0, metric="cosine",
random_state=s * 7 + 1, low_memory=True)
red = u.fit_transform(embeddings)
h = HDBSCAN(min_cluster_size=params["min_cluster_size"],
min_samples=params["min_samples"],
metric="euclidean",
cluster_selection_method=params["csm"],
cluster_selection_epsilon=params["cse"])
all_labels.append(h.fit_predict(red))
aris = []
for i in range(len(all_labels)):
for j in range(i + 1, len(all_labels)):
aris.append(adjusted_rand_score(all_labels[i], all_labels[j]))
return float(np.mean(aris)) if aris else 0.0
# ---------------------------------------------------------------------------
# §3.4 — Bayesian optimisation objective
# ---------------------------------------------------------------------------
def _objective(trial, embeddings, n_docs):
"""Single Optuna trial — returns (persistence, dbcv, stability_placeholder)."""
n_neighbors = trial.suggest_categorical("n_neighbors", [5, 10, 15, 30, 50])
n_components = trial.suggest_int("n_components", 5, 10)
mcs = trial.suggest_int(
"min_cluster_size",
max(5, int(0.01 * n_docs)),
max(20, int(0.05 * n_docs)),
)
ms = trial.suggest_int("min_samples", 1, mcs)
csm = trial.suggest_categorical("csm", ["eom", "leaf"])
cse = trial.suggest_float("cse", 0.0, 0.3, step=0.05)
params = dict(n_neighbors=n_neighbors, n_components=n_components,
min_cluster_size=mcs, min_samples=ms, csm=csm, cse=cse)
u = UMAP(n_neighbors=n_neighbors, n_components=n_components,
min_dist=0.0, metric="cosine", random_state=42,
low_memory=True)
red = u.fit_transform(embeddings)
h = HDBSCAN(min_cluster_size=mcs, min_samples=ms, metric="euclidean",
cluster_selection_method=csm,
cluster_selection_epsilon=cse,
allow_single_cluster=False, gen_min_span_tree=True)
labels = h.fit_predict(red)
disc = check_discipline(labels, n_docs)
trial.set_user_attr("params", params)
trial.set_user_attr("discipline", disc)
trial.set_user_attr("labels", labels.tolist())
# Hard-constraint violation → worst scores
if not disc["max_mass_ok"] or not disc["min_size_ok"]:
trial.set_user_attr("pass", False)
return 0.0, -1.0, 0.0
trial.set_user_attr("pass", True)
pers = compute_persistence(h)
dbcv = compute_dbcv(red, labels)
trial.set_user_attr("persistence", pers)
trial.set_user_attr("dbcv", dbcv)
return pers, dbcv, 0.5 # stability computed only for winner
# ---------------------------------------------------------------------------
# §3.4 — Run the full Bayesian loop
# ---------------------------------------------------------------------------
def run_bayesian_optimisation(
embeddings: np.ndarray,
n_trials: int = 50,
progress_callback=None,
) -> dict:
n_docs = len(embeddings)
study = optuna.create_study(
directions=["maximize", "maximize", "maximize"],
sampler=optuna.samplers.TPESampler(seed=42, multivariate=True),
study_name="specter2_umap_hdbscan",
)
trial_log = []
def _cb(study, trial):
d = trial.user_attrs.get("discipline", {})
entry = dict(
trial=trial.number,
params=trial.user_attrs.get("params", {}),
discipline_pass=trial.user_attrs.get("pass", False),
persistence=trial.user_attrs.get("persistence", 0.0),
dbcv=trial.user_attrs.get("dbcv", -1.0),
n_clusters=d.get("n_clusters", 0),
max_mass_pct=d.get("max_mass_pct", 0.0),
min_size=d.get("min_size", 0),
n_noise=d.get("n_noise", 0),
values=list(trial.values) if trial.values else [],
)
trial_log.append(entry)
if progress_callback:
progress_callback(trial.number + 1, n_trials, entry)
for i in range(n_trials):
study.optimize(
lambda t: _objective(t, embeddings, n_docs),
n_trials=1, callbacks=[_cb], show_progress_bar=False,
)
# §3.6 convergence: 3 consecutive passing within 5 % of best
passing = [e for e in trial_log if e["discipline_pass"]]
if len(passing) >= 3 and i >= 9: # allow early stop after 10 trials
best_p = max(e["persistence"] for e in passing)
if best_p > 0:
last3 = passing[-3:]
if all(abs(e["persistence"] - best_p) / best_p < 0.05
for e in last3):
logger.info("Converged at trial %d.", i + 1)
break
# Select best passing trial (max persistence, then DBCV)
passing_trials = [t for t in study.trials
if t.user_attrs.get("pass", False)]
if passing_trials:
best = max(passing_trials, key=lambda t: (t.values[0], t.values[1]))
else:
logger.warning("No trial passed discipline — using last trial.")
best = study.trials[-1]
bp = best.user_attrs["params"]
labels = np.array(best.user_attrs["labels"])
stability = compute_stability(embeddings, bp, n_seeds=3)
return dict(
best_params=bp, best_labels=labels,
best_trial=best.number,
persistence=best.user_attrs.get("persistence", 0.0),
dbcv=best.user_attrs.get("dbcv", -1.0),
stability=stability,
discipline=best.user_attrs.get("discipline", {}),
trial_log=trial_log,
n_trials_run=len(trial_log),
)
# ---------------------------------------------------------------------------
# §3.1 — 2-D UMAP for visualisation
# ---------------------------------------------------------------------------
def compute_2d_umap(embeddings: np.ndarray, seed: int = 42) -> np.ndarray:
return UMAP(n_neighbors=15, n_components=2, min_dist=0.1,
metric="cosine", random_state=seed,
low_memory=True).fit_transform(embeddings)
# ---------------------------------------------------------------------------
# §3.1 — TF-IDF keyphrase extraction per cluster (3–5 phrases)
# Fast alternative to KeyBERT — no extra model download needed.
# ---------------------------------------------------------------------------
def extract_keyphrases(docs: list[str], labels: np.ndarray,
top_n: int = 5) -> dict:
from sklearn.feature_extraction.text import TfidfVectorizer
cluster_docs = defaultdict(list)
for doc, lab in zip(docs, labels):
if lab != -1:
cluster_docs[int(lab)].append(doc)
out = {}
for cid, cdocs in cluster_docs.items():
if len(cdocs) < 2:
out[cid] = []
continue
try:
tfidf = TfidfVectorizer(
stop_words="english", max_features=200,
ngram_range=(1, 3), max_df=0.9, min_df=1)
X = tfidf.fit_transform(cdocs)
terms = tfidf.get_feature_names_out()
scores = X.sum(axis=0).A1
top_idx = scores.argsort()[::-1][:top_n]
out[cid] = [(terms[i], float(scores[i])) for i in top_idx]
except Exception as e:
logger.warning("Keyphrase extraction cluster %d: %s", cid, e)
out[cid] = []
return out
# ---------------------------------------------------------------------------
# §3.1 — Strong / weak member counts via HDBSCAN probabilities
# ---------------------------------------------------------------------------
def strong_weak_members(labels: np.ndarray,
probabilities: np.ndarray) -> dict:
mem = defaultdict(lambda: {"strong": 0, "weak": 0})
for lab, prob in zip(labels, probabilities):
if lab == -1:
continue
cid = int(lab)
if prob >= 0.5:
mem[cid]["strong"] += 1
else:
mem[cid]["weak"] += 1
return dict(mem)
# ---------------------------------------------------------------------------
# §3.2 — Outlier reduction: reassign noise to nearest cluster (≤ 25 %)
# ---------------------------------------------------------------------------
def outlier_reduction(labels: np.ndarray, reduced: np.ndarray,
n_docs: int) -> np.ndarray:
labels = labels.copy()
cap = int(0.25 * n_docs)
cdocs = defaultdict(list)
for i, l in enumerate(labels):
if l != -1:
cdocs[int(l)].append(i)
if not cdocs:
return labels
cids = list(cdocs.keys())
centroids = np.vstack([reduced[cdocs[c]].mean(axis=0) for c in cids])
noise = [i for i, l in enumerate(labels) if l == -1]
moved = 0
for idx in noise:
dists = np.linalg.norm(centroids - reduced[idx], axis=1)
for best in np.argsort(dists):
tgt = cids[best]
if len(cdocs[tgt]) < cap:
labels[idx] = tgt
cdocs[tgt].append(idx)
moved += 1
break
logger.info("Outlier reduction: %d / %d noise reassigned.", moved, len(noise))
return labels
# ---------------------------------------------------------------------------
# Representative docs (top-3 by centroid proximity)
# ---------------------------------------------------------------------------
def get_representative_docs(labels, embeddings, docs, top_n=3):
cdocs = defaultdict(list)
for i, l in enumerate(labels):
if l != -1:
cdocs[int(l)].append(i)
out = {}
for cid, idxs in cdocs.items():
ce = embeddings[idxs].mean(axis=0).reshape(1, -1)
sims = cosine_similarity(ce, embeddings[idxs])[0]
top = np.argsort(sims)[-top_n:][::-1]
out[cid] = [docs[idxs[t]] for t in top]
return out
# ---------------------------------------------------------------------------
# §9 — RQ2 / RQ3 mismatch table
# ---------------------------------------------------------------------------
def build_mismatch_table(keyphrases: dict, cluster_labels: dict) -> list:
"""Compare cluster keyphrases against assigned labels to flag mismatches.
Returns rows for a mismatch table (§9)."""
rows = []
for cid in sorted(keyphrases.keys()):
kps = keyphrases.get(cid, [])
kp_terms = [k[0] if isinstance(k, tuple) else k for k in kps[:5]]
label = cluster_labels.get(cid, "")
# Check overlap between label words and keyphrase terms
label_words = set(label.lower().split())
kp_words = set(" ".join(kp_terms).lower().split())
overlap = label_words & kp_words
noise = {"the","and","for","with","using","based","from","in","of","a","to"}
overlap -= noise
match_pct = len(overlap) / max(len(label_words - noise), 1)
status = "MATCH" if match_pct >= 0.3 else "MISMATCH"
rows.append(dict(
cluster=cid, label=label,
keyphrases=", ".join(kp_terms),
overlap=", ".join(overlap) if overlap else "—",
match_pct=round(match_pct * 100),
status=status,
))
return rows
# ---------------------------------------------------------------------------
# High-level pipeline entry point
# ---------------------------------------------------------------------------
def run_topic_modeling(filepath: str, n_trials: int = 50,
progress_callback=None) -> dict:
# 1. Load
df = load_csv(filepath)
docs = prepare_documents(df)
n_docs = len(docs)
# 2. Embed (deterministic)
embeddings = embed_documents(docs)
# 3. Bayesian optimisation (§3.4)
opt = run_bayesian_optimisation(embeddings, n_trials, progress_callback)
bp = opt["best_params"]
labels = opt["best_labels"]
# 4. Re-run winner for clusterer object (probabilities)
u = UMAP(n_neighbors=bp["n_neighbors"], n_components=bp["n_components"],
min_dist=0.0, metric="cosine", random_state=42)
red = u.fit_transform(embeddings)
h = HDBSCAN(min_cluster_size=bp["min_cluster_size"],
min_samples=bp["min_samples"], metric="euclidean",
cluster_selection_method=bp["csm"],
cluster_selection_epsilon=bp["cse"],
allow_single_cluster=False,
gen_min_span_tree=True)
h.fit(red)
# Per-cluster persistence (§8)
cluster_pers = per_cluster_persistence(h, labels)
# 5. Outlier reduction (§3.2 — clusters < 5 reassigned)
labels = outlier_reduction(labels, red, n_docs)
# 6. Strong / weak (§3.1)
sw = strong_weak_members(labels, h.probabilities_)
# 7. 2-D UMAP (§3.1)
umap_2d = compute_2d_umap(embeddings)
# 8. KeyBERT keyphrases (§3.1)
keyphrases = extract_keyphrases(docs, labels)
# 9. Rep docs
rep_docs = get_representative_docs(labels, embeddings, docs)
# 10. Final discipline
disc = check_discipline(labels, n_docs)
return dict(
documents=docs, labels=labels.tolist(),
keyphrases=keyphrases, representative_docs=rep_docs,
membership=sw, umap_2d=umap_2d.tolist(),
discipline=disc, best_params=bp,
cluster_persistence=cluster_pers,
metrics=dict(persistence=opt["persistence"],
dbcv=opt["dbcv"],
stability=opt["stability"]),
trial_log=opt["trial_log"],
n_trials_run=opt["n_trials_run"],
best_trial=opt["best_trial"],
n_docs=n_docs,
embeddings=embeddings,
)
|