File size: 10,508 Bytes
8dcc0e4 | 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 | """Stage 3: Feature engineering for fraud detection."""
from __future__ import annotations
import json
from pathlib import Path
import sys
from typing import Iterable
import numpy as np
import pandas as pd
if __package__ in (None, ""):
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from src.config import load_params
def compute_feature_thresholds(df: pd.DataFrame) -> dict[str, float]:
"""Compute training-time thresholds used by percentile / z-score features.
These values MUST be persisted and reused at inference time — computing
them from a single inference row produces degenerate features (e.g.,
`is_p95_amount` becomes 0 for every row since `amount.quantile(0.95)`
equals the row's own amount).
Parameters
----------
df : pd.DataFrame
Standardized transaction dataframe containing an `amount` column.
Returns
-------
dict[str, float]
Thresholds: amount_p95, amount_p99, amount_mean, amount_std.
"""
amount = pd.to_numeric(df["amount"], errors="coerce").fillna(0.0)
return {
"amount_p95": float(amount.quantile(0.95)) if len(amount) > 0 else 200000.0,
"amount_p99": float(amount.quantile(0.99)) if len(amount) > 0 else 500000.0,
"amount_mean": float(amount.mean()) if len(amount) > 0 else 0.0,
"amount_std": float(amount.std()) if len(amount) > 1 else 1.0,
}
def _add_amount_bin_partition(series: pd.Series) -> pd.Series:
"""Create amount bins in a pandas partition.
Parameters
----------
series : pd.Series
Amount series.
Returns
-------
pd.Series
Categorical bin labels.
"""
bins = [-np.inf, 1000, 10000, 100000, np.inf]
labels = ["low", "medium", "high", "very_high"]
return pd.cut(series, bins=bins, labels=labels).astype("string")
def engineer_features_df(df: pd.DataFrame) -> pd.DataFrame:
"""Apply feature engineering using pandas operations.
Parameters
----------
df : pd.DataFrame
Combined standardized transaction dataset.
Returns
-------
pd.DataFrame
Engineered feature dataframe with encoded categoricals.
"""
required = {
"amount",
"is_fraud",
"transaction_type",
"hour_of_day",
"day_of_month",
"balance_change_orig",
"balance_ratio",
"has_balance_info",
"source",
"txn_timestamp",
}
missing = required - set(df.columns)
if missing:
raise KeyError(f"[Preprocess] Missing required columns: {sorted(missing)}")
amount = pd.to_numeric(df["amount"], errors="coerce").fillna(0.0)
hour = pd.to_numeric(df["hour_of_day"], errors="coerce").fillna(0).astype("int64")
day = pd.to_numeric(df["day_of_month"], errors="coerce").fillna(1).astype("int64")
ts = pd.to_numeric(df["txn_timestamp"], errors="coerce").fillna(0).astype("int64")
bal_change = pd.to_numeric(df["balance_change_orig"], errors="coerce")
bal_ratio = pd.to_numeric(df["balance_ratio"], errors="coerce")
bal_change_abs = bal_change.abs()
median_change = float(bal_change_abs.quantile(0.5))
bal_change_abs = bal_change_abs.fillna(median_change)
bal_ratio_clipped = bal_ratio.fillna(0.0).clip(-10, 10)
# Percentile-based amount flags
amount_p95 = float(amount.quantile(0.95)) if len(amount) > 0 else 200000
amount_p99 = float(amount.quantile(0.99)) if len(amount) > 0 else 500000
out = df.assign(
amount=amount,
is_fraud=pd.to_numeric(df["is_fraud"], errors="coerce").fillna(0).astype("int64"),
txn_timestamp=ts,
tx_amount_log=np.log1p(amount),
is_high_amount=(amount > 200000).astype("int64"),
is_p95_amount=(amount > amount_p95).astype("int64"),
is_p99_amount=(amount > amount_p99).astype("int64"),
amount_zscore=((amount - amount.mean()) / (amount.std() + 1e-8)),
hour_sin=np.sin(2.0 * np.pi * hour / 24.0),
hour_cos=np.cos(2.0 * np.pi * hour / 24.0),
day_sin=np.sin(2.0 * np.pi * (day % 31) / 31.0),
day_cos=np.cos(2.0 * np.pi * (day % 31) / 31.0),
is_night=(((hour >= 0) & (hour < 6)) | (hour >= 22)).astype("int64"),
is_weekend=((day % 7).isin([5, 6])).astype("int64"),
balance_change_abs=bal_change_abs,
balance_ratio_clipped=bal_ratio_clipped,
balance_change_log=np.log1p(bal_change_abs),
)
out["amount_bin"] = _add_amount_bin_partition(out["amount"])
out = pd.get_dummies(out, columns=["transaction_type", "source", "amount_bin"], dtype="int64")
return out
def engineer_features_pandas(
df: pd.DataFrame,
thresholds: dict[str, float] | None = None,
) -> pd.DataFrame:
"""Apply the same feature engineering logic in pandas for inference.
Parameters
----------
df : pd.DataFrame
Input standardized dataframe.
thresholds : dict[str, float] | None
Training-time thresholds (amount_p95, amount_p99, amount_mean,
amount_std). When provided, these are used instead of recomputing
from `df` — which is essential for single-row or small-batch
inference where inference-data statistics are meaningless.
Returns
-------
pd.DataFrame
Engineered dataframe.
"""
def _series_or_default(column: str, default: float | int) -> pd.Series:
if column in df.columns:
return pd.to_numeric(df[column], errors="coerce")
return pd.Series(default, index=df.index)
amount = _series_or_default("amount", 0.0).fillna(0.0)
hour = _series_or_default("hour_of_day", 0).fillna(0).astype(int)
day = _series_or_default("day_of_month", 1).fillna(1).astype(int)
bal_change = _series_or_default("balance_change_orig", 0.0)
bal_ratio = _series_or_default("balance_ratio", 0.0)
bal_change_abs = bal_change.abs()
med = float(bal_change_abs.dropna().median()) if bal_change_abs.notna().any() else 0.0
if thresholds:
# Prefer training-time thresholds to avoid leakage from inference data.
amount_p95 = float(thresholds.get("amount_p95", 200000.0))
amount_p99 = float(thresholds.get("amount_p99", 500000.0))
amount_mean = float(thresholds.get("amount_mean", 0.0))
amount_std = float(thresholds.get("amount_std", 1.0))
else:
# Fallback: recompute from input (degenerate for single-row inference).
amount_p95 = float(amount.quantile(0.95)) if len(amount) > 1 else 200000.0
amount_p99 = float(amount.quantile(0.99)) if len(amount) > 1 else 500000.0
amount_mean = float(amount.mean()) if len(amount) > 0 else 0.0
amount_std = float(amount.std()) if len(amount) > 1 else 1.0
out = df.copy()
out["amount"] = amount
out["tx_amount_log"] = np.log1p(amount)
out["is_high_amount"] = (amount > 200000).astype(int)
out["is_p95_amount"] = (amount > amount_p95).astype(int)
out["is_p99_amount"] = (amount > amount_p99).astype(int)
out["amount_zscore"] = (amount - amount_mean) / (amount_std + 1e-8)
out["hour_sin"] = np.sin(2.0 * np.pi * hour / 24.0)
out["hour_cos"] = np.cos(2.0 * np.pi * hour / 24.0)
out["day_sin"] = np.sin(2.0 * np.pi * (day % 31) / 31.0)
out["day_cos"] = np.cos(2.0 * np.pi * (day % 31) / 31.0)
out["is_night"] = (((hour >= 0) & (hour < 6)) | (hour >= 22)).astype(int)
out["is_weekend"] = ((day % 7).isin([5, 6])).astype(int)
out["balance_change_abs"] = bal_change_abs.fillna(med)
out["balance_ratio_clipped"] = bal_ratio.fillna(0.0).clip(-10, 10)
out["balance_change_log"] = np.log1p(bal_change_abs.fillna(med))
out["amount_bin"] = pd.cut(
out["amount"],
bins=[-np.inf, 1000, 10000, 100000, np.inf],
labels=["low", "medium", "high", "very_high"],
).astype("string")
for col in ("transaction_type", "source"):
if col not in out.columns:
out[col] = "unknown"
out = pd.get_dummies(out, columns=["transaction_type", "source", "amount_bin"], dtype=int)
return out
def align_feature_columns(df: pd.DataFrame, feature_columns: Iterable[str]) -> pd.DataFrame:
"""Align a dataframe to model feature columns.
Parameters
----------
df : pd.DataFrame
Feature dataframe.
feature_columns : Iterable[str]
Expected feature column order.
Returns
-------
pd.DataFrame
Aligned dataframe.
"""
aligned = df.copy()
for col in feature_columns:
if col not in aligned.columns:
aligned[col] = 0
extra = [c for c in aligned.columns if c not in feature_columns]
if extra:
aligned = aligned.drop(columns=extra)
return aligned[list(feature_columns)]
def main() -> None:
"""Run Stage 3 pipeline step."""
params = load_params()
cfg = params.get("preprocess", {})
input_path = cfg.get("input", "data/processed/combined_transactions.csv")
output_path = cfg.get("output", "data/processed/features.csv")
if not Path(input_path).exists():
raise FileNotFoundError(f"[Preprocess] Input file not found: {input_path}")
print(f"[Preprocess] Loading {input_path}")
df = pd.read_csv(input_path)
thresholds = compute_feature_thresholds(df)
features_df = engineer_features_df(df)
Path(output_path).parent.mkdir(parents=True, exist_ok=True)
features_df.to_csv(output_path, index=False)
thresholds_path = Path(output_path).parent / "feature_thresholds.json"
with open(thresholds_path, "w", encoding="utf-8") as f:
json.dump(thresholds, f, indent=2)
feature_names = list(features_df.columns)
shape_rows = int(features_df.shape[0])
null_counts = features_df.isnull().sum().to_dict()
fraud_dist = features_df["is_fraud"].value_counts().to_dict()
print(f"[Preprocess] Saved features to {output_path}")
print(f"[Preprocess] Saved thresholds to {thresholds_path}: {thresholds}")
print(f"[Preprocess] Shape: ({shape_rows}, {len(feature_names)})")
print(f"[Preprocess] Feature names: {feature_names}")
print(f"[Preprocess] Null counts: {null_counts}")
print(f"[Preprocess] Fraud distribution: {fraud_dist}")
if __name__ == "__main__":
main()
|