File size: 12,341 Bytes
66242b8 | 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 | from __future__ import annotations
import json
from xgboost import XGBClassifier
from helpers import save_json
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any
import numpy as np
import pandas as pd
from scipy import sparse
from sklearn.metrics import (
accuracy_score,
balanced_accuracy_score,
confusion_matrix,
f1_score,
precision_score,
recall_score,
roc_auc_score,
)
@dataclass(slots=True)
class Config:
# local pairwise
include_statistical: bool = True
include_tfidf: bool = True
include_char_ngrams: bool = True
include_pos_ngrams: bool = True
include_readability: bool = True
include_local_pairwise: bool = True # local features pairwise operations; this will override the local pairwise booleans config
include_global_pairwise: bool = True # global features using cosine similarity
threshold_metric: str = "youden_j" # reducing false positives
threshold_grid_step: float = 0.01
model_params: dict[str, Any] | None = None
def __post_init__(self) -> None:
if self.model_params is None: # hyperparameters for XGBoost
# include local features
self.model_params = {
"objective": "binary:logistic",
"eval_metric": "logloss",
"n_estimators": 500,
"max_depth": 4,
"learning_rate": 0.05,
"subsample": 0.8,
"colsample_bytree": 0.3,
"min_child_weight": 3,
"reg_lambda": 5.0,
"reg_alpha": 1.0,
"random_state": 42,
"n_jobs": 2,
"tree_method": "hist",
}
# # only global featuers
# self.model_params = {
# "objective": "binary:logistic",
# "eval_metric": "logloss",
# "n_estimators": 200,
# "max_depth": 2,
# "learning_rate": 0.03,
# "subsample": 0.9,
# "colsample_bytree": 1.0,
# "min_child_weight": 5,
# "reg_lambda": 10.0,
# "reg_alpha": 0.5,
# "random_state": 42,
# "n_jobs": 2,
# "tree_method": "hist",
# }
def _feature_family_from_suffix(suffix: str) -> str:
if suffix.startswith("tfidf_"):
return "tfidf"
if suffix.startswith("char") and "_tfidf_" in suffix:
return "char_ngrams"
if suffix.startswith("pos") and "_tfidf_" in suffix:
return "pos_ngrams"
if suffix.startswith("readability_"):
return "readability"
return "statistical"
def _include_family(family: str, config: Config) -> bool:
return {
"statistical": config.include_statistical,
"tfidf": config.include_tfidf,
"char_ngrams": config.include_char_ngrams,
"pos_ngrams": config.include_pos_ngrams,
"readability": config.include_readability,
}[family]
# self-build cosine similarity function
def _safe_cosine_from_columns(left: sparse.csr_matrix, right: sparse.csr_matrix) -> np.ndarray:
numerator = np.asarray(left.multiply(right).sum(axis=1)).ravel()
left_norm = np.sqrt(np.asarray(left.multiply(left).sum(axis=1)).ravel())
right_norm = np.sqrt(np.asarray(right.multiply(right).sum(axis=1)).ravel())
denominator = left_norm * right_norm
result = np.divide(numerator, denominator, out=np.zeros_like(numerator, dtype=np.float32), where=denominator > 0)
return result.astype(np.float32)
def discover_suffixes(train_df: pd.DataFrame, config: Config) -> list[str]:
suffixes: list[str] = []
for column in train_df.columns:
if not column.startswith("text1_"):
continue
suffix = column[len("text1_") :]
if _include_family(_feature_family_from_suffix(suffix), config):
suffixes.append(suffix)
return suffixes
# summary global features
def build_global_pairwise_features(df: pd.DataFrame, suffixes: list[str]) -> tuple[Any, list[str]]:
dtype = np.float32
blocks = []
feature_names = []
families = {
"tfidf": [s for s in suffixes if s.startswith("tfidf_")],
"char_ngrams": [s for s in suffixes if s.startswith("char") and "_tfidf_" in s],
"pos_ngrams": [s for s in suffixes if s.startswith("pos") and "_tfidf_" in s],
"scalar": [s for s in suffixes if not (
s.startswith("tfidf_")
or (s.startswith("char") and "_tfidf_" in s)
or (s.startswith("pos") and "_tfidf_" in s)
)
]}
for family_name, family_suffixes in families.items():
if not family_suffixes:
continue
left_cols = [f"text1_{s}" for s in family_suffixes]
right_cols = [f"text2_{s}" for s in family_suffixes]
left = sparse.csr_matrix(df[left_cols].to_numpy(dtype=dtype))
right = sparse.csr_matrix(df[right_cols].to_numpy(dtype=dtype))
diff = left - right
cosine = _safe_cosine_from_columns(left, right).reshape(-1, 1) # cosine similarity
l1 = np.asarray(np.abs(diff).sum(axis=1)).ravel().astype(dtype).reshape(-1, 1) # l1 distance
l2 = np.sqrt(np.asarray(diff.multiply(diff).sum(axis=1)).ravel()).astype(dtype).reshape(-1, 1) # l2 distance
family_block = sparse.csr_matrix(np.hstack([cosine, l1, l2]), dtype=dtype)
blocks.append(family_block)
feature_names.extend([
f"{family_name}_cosine_similarity",
f"{family_name}_l1_distance",
f"{family_name}_l2_distance",
])
if not blocks:
return sparse.csr_matrix((len(df), 0), dtype=dtype), []
return sparse.hstack(blocks, format="csr", dtype=dtype), feature_names
# local features (pairwise operations done for each feature column)
def build_pairwise_matrix(df: pd.DataFrame, suffixes: list[str]) -> tuple[Any, np.ndarray, list[str]]:
dtype = np.float32
feature_names: list[str] = []
columns: list[sparse.csr_matrix] = []
for suffix in suffixes:
# operate one for each column for pairwise operations
left = sparse.csr_matrix(df[f"text1_{suffix}"].to_numpy(dtype=dtype).reshape(-1, 1))
right = sparse.csr_matrix(df[f"text2_{suffix}"].to_numpy(dtype=dtype).reshape(-1, 1))
diff = left - right
# asbolute difference
columns.append(abs(diff))
feature_names.append(f"{suffix}_abs_diff")
# dot product
columns.append(left.multiply(right))
feature_names.append(f"{suffix}_product")
X = sparse.hstack(columns, format="csr", dtype=dtype)
y = df["same"].to_numpy(dtype=np.int8, copy=False) # binary label
return X, y, feature_names
def compute_metrics(y_true: np.ndarray, y_proba: np.ndarray, threshold: float) -> dict[str, Any]:
y_pred = (y_proba >= threshold).astype(int)
tn, fp, fn, tp = confusion_matrix(y_true, y_pred, labels=[0, 1]).ravel()
specificity = tn / (tn + fp) if (tn + fp) > 0 else 0.0
sensitivity = recall_score(y_true, y_pred, zero_division=0)
balanced_accuracy = balanced_accuracy_score(y_true, y_pred)
youden_j = sensitivity + specificity - 1.0
return {
"threshold": round(threshold, 5),
"accuracy": round(accuracy_score(y_true, y_pred), 5),
"precision": round(precision_score(y_true, y_pred, zero_division=0), 5),
"recall": round(sensitivity, 5),
"f1": round(f1_score(y_true, y_pred, zero_division=0), 5),
"balanced_accuracy": round(balanced_accuracy, 5),
"specificity": round(specificity, 5),
"youden_j": round(youden_j, 5),
"roc_auc": round(roc_auc_score(y_true, y_proba), 5),
"tn": int(tn),
"fp": int(fp),
"fn": int(fn),
"tp": int(tp),
}
# finding best threshold using grid search based on config.threshold_metric
# using different config.threshold_metric can lead to different performance for the classification (not proba)
def find_best_threshold(
y_true: np.ndarray,
y_proba: np.ndarray,
config: Config,
) -> tuple[float, dict[str, Any]]:
thresholds = np.arange(0.0, 1.0+config.threshold_grid_step, config.threshold_grid_step, dtype=np.float32)
if thresholds.size == 0:
thresholds = np.array([0.5], dtype=np.float32)
best_threshold = 0.5
best_metrics = compute_metrics(y_true, y_proba, threshold=best_threshold)
best_score = float(best_metrics[config.threshold_metric])
for threshold in thresholds:
metrics = compute_metrics(y_true, y_proba, threshold=float(threshold)) # compute metrics' value for each test threshold
score = float(metrics[config.threshold_metric])
# if current score > best_score...
if score > best_score:
best_threshold = float(threshold)
best_metrics = metrics
best_score = score
return best_threshold, best_metrics
def train_and_save_model(save_root: str | Path | None = None, config: Config | None = None) -> dict[str, Any]:
config = config or Config()
project_root = Path(__file__).resolve().parents[1] # assuming this file under subfolder in project root
saved_dir = project_root / "saved"
ngram_dir = saved_dir / "ngram_features" / "dataframes"
model_dir = Path(save_root) if save_root is not None else saved_dir / "model"
ngram_dict_df = {split: pd.read_parquet(ngram_dir / f"{split}_ngram.parquet") for split in ("train", "validation", "test")}
suffixes = discover_suffixes(ngram_dict_df["train"], config)
X_by_split: dict[str, Any] = {}
y_by_split: dict[str, np.ndarray] = {}
feature_names: list[str] = []
for split, df in ngram_dict_df.items():
blocks = []
feature_names = []
if config.include_local_pairwise:
X_pairwise, y, local_names = build_pairwise_matrix(df, suffixes)
blocks.append(X_pairwise)
feature_names.extend(local_names)
else:
y = df["same"].to_numpy(dtype=np.int8, copy=False)
if config.include_global_pairwise:
X_global, global_names = build_global_pairwise_features(df, suffixes)
blocks.append(X_global)
feature_names.extend(global_names)
if not blocks: # if both local and global features are set to false
raise ValueError("At least one of include_local_pairwise or include_global_pairwise must be True.")
X = sparse.hstack(blocks, format="csr", dtype=np.float32)
X_by_split[split] = X
y_by_split[split] = y
model = XGBClassifier(**config.model_params) # the model can be changed as desired
model.fit(X_by_split["train"], y_by_split["train"]) # fitting
validation_proba = model.predict_proba(X_by_split["validation"])[:, 1]
best_threshold, validation_metrics = find_best_threshold(
y_by_split["validation"],
validation_proba,
config)
test_proba = model.predict_proba(X_by_split["test"])[:, 1]
test_metrics = compute_metrics(y_by_split["test"], test_proba, threshold=best_threshold)
# saving
model_dir.mkdir(parents=True, exist_ok=True)
model.save_model(model_dir / "model.json")
save_json({"threshold": best_threshold}, model_dir / "threshold.json") # saving best threshold
save_json({ # saving features used for training, including original features
"suffixes": suffixes,
"feature_names": feature_names,
},
model_dir / "feature_spec.json")
save_json({ # saving model performance metrics
"validation": validation_metrics,
"test": test_metrics,
},
model_dir / "metrics.json")
save_json(asdict(config), model_dir / "training_config.json") # saving model config
return {
"model_dir": str(model_dir),
"model_path": str(model_dir / "model.json"),
"threshold": best_threshold,
"suffixes": suffixes,
"feature_names": feature_names,
"metrics": {
"validation": validation_metrics,
"test": test_metrics,
}}
if __name__ == "__main__":
outputs = train_and_save_model()
print(f"Saved model bundle to: {outputs['model_dir']}")
print(json.dumps(outputs["metrics"], indent=2))
|