File size: 6,044 Bytes
572c4ce | 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 | from __future__ import annotations
import json
from pathlib import Path
import joblib
import numpy as np
import pandas as pd
import trackio
from sklearn.ensemble import HistGradientBoostingClassifier, IsolationForest
from sklearn.metrics import (
average_precision_score,
confusion_matrix,
f1_score,
precision_score,
recall_score,
roc_auc_score,
)
from sklearn.preprocessing import RobustScaler
PROJECT_DIR = Path(__file__).resolve().parent
DATA_DIR = PROJECT_DIR / "data"
ARTIFACT_DIR = PROJECT_DIR / "artifacts" / "edge-sentinel-ensemble"
EXCLUDED = {"device_id", "time", "label", "anomaly_type", "split"}
def load_split(name: str) -> pd.DataFrame:
return pd.read_parquet(DATA_DIR / f"{name}.parquet")
def metrics(labels: np.ndarray, scores: np.ndarray, threshold: float) -> dict:
predictions = (scores >= threshold).astype(np.int8)
matrix = confusion_matrix(labels, predictions, labels=[0, 1])
return {
"roc_auc": float(roc_auc_score(labels, scores)),
"average_precision": float(average_precision_score(labels, scores)),
"precision": float(precision_score(labels, predictions, zero_division=0)),
"recall": float(recall_score(labels, predictions, zero_division=0)),
"f1": float(f1_score(labels, predictions, zero_division=0)),
"false_positive_rate": float(matrix[0, 1] / max(1, matrix[0].sum())),
"confusion_matrix": matrix.tolist(),
}
def best_threshold(labels: np.ndarray, scores: np.ndarray) -> tuple[float, float]:
candidates = np.quantile(scores, np.linspace(0.7, 0.999, 300))
best = (0.5, -1.0)
for threshold in candidates:
value = f1_score(labels, scores >= threshold, zero_division=0)
if value > best[1]:
best = (float(threshold), float(value))
return best
def normalize(values: np.ndarray, reference: np.ndarray) -> np.ndarray:
low, high = np.quantile(reference, [0.01, 0.99])
return np.clip((values - low) / max(1e-9, high - low), 0, 1)
def main() -> None:
train = load_split("train")
validation = load_split("validation")
test = load_split("test")
features = [column for column in train.columns if column not in EXCLUDED]
scaler = RobustScaler()
x_train = scaler.fit_transform(train[features])
x_validation = scaler.transform(validation[features])
x_test = scaler.transform(test[features])
y_train = train["label"].to_numpy()
y_validation = validation["label"].to_numpy()
y_test = test["label"].to_numpy()
isolation = IsolationForest(
n_estimators=300,
max_samples=0.7,
contamination="auto",
random_state=42,
n_jobs=-1,
)
isolation.fit(x_train[y_train == 0])
supervised = HistGradientBoostingClassifier(
learning_rate=0.08,
max_iter=250,
max_leaf_nodes=31,
l2_regularization=0.2,
class_weight="balanced",
random_state=42,
)
supervised.fit(x_train, y_train)
isolation_validation_raw = -isolation.score_samples(x_validation)
isolation_test_raw = -isolation.score_samples(x_test)
isolation_validation = normalize(isolation_validation_raw, isolation_validation_raw)
isolation_test = normalize(isolation_test_raw, isolation_validation_raw)
supervised_validation = supervised.predict_proba(x_validation)[:, 1]
supervised_test = supervised.predict_proba(x_test)[:, 1]
trackio.init(
project="edge-sentinel-ml",
name="device-held-out-ensemble-v1",
config={
"train_devices": 8,
"validation_devices": 2,
"test_devices": 2,
"features": len(features),
},
)
candidates = []
for supervised_weight in np.linspace(0, 1, 21):
validation_scores = (
supervised_weight * supervised_validation
+ (1 - supervised_weight) * isolation_validation
)
threshold, validation_f1 = best_threshold(y_validation, validation_scores)
candidates.append((validation_f1, supervised_weight, threshold))
trackio.log(
{
"supervised_weight": supervised_weight,
"validation_f1": validation_f1,
"threshold": threshold,
}
)
_, weight, threshold = max(candidates)
validation_scores = weight * supervised_validation + (1 - weight) * isolation_validation
test_scores = weight * supervised_test + (1 - weight) * isolation_test
results = {
"features": features,
"ensemble_supervised_weight": float(weight),
"threshold": threshold,
"validation": metrics(y_validation, validation_scores, threshold),
"test": metrics(y_test, test_scores, threshold),
"isolation_test": metrics(
y_test,
isolation_test,
best_threshold(y_validation, isolation_validation)[0],
),
"supervised_test": metrics(
y_test,
supervised_test,
best_threshold(y_validation, supervised_validation)[0],
),
}
trackio.log(
{
"test_roc_auc": results["test"]["roc_auc"],
"test_average_precision": results["test"]["average_precision"],
"test_f1": results["test"]["f1"],
"test_recall": results["test"]["recall"],
"test_false_positive_rate": results["test"]["false_positive_rate"],
}
)
trackio.finish()
ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
joblib.dump(
{
"scaler": scaler,
"isolation_forest": isolation,
"supervised_model": supervised,
"features": features,
"weight": weight,
"threshold": threshold,
},
ARTIFACT_DIR / "edge_sentinel.joblib",
compress=3,
)
(ARTIFACT_DIR / "evaluation.json").write_text(
json.dumps(results, indent=2),
encoding="utf-8",
)
print(json.dumps(results, indent=2))
if __name__ == "__main__":
main()
|