cyb008-baseline-classifier / feature_engineering.py
pradeep-xpert's picture
Initial release: XGBoost + MLP for SOC alert triage outcome classification, with structural-leakage and unlearnable-target diagnostic
001717c verified
"""
feature_engineering.py
======================
Feature pipeline for the CYB008 baseline classifier.
Predicts `resolution_outcome` (5-class triage outcome) from per-alert
features on the CYB008 sample dataset.
CSV inputs:
soc_alerts.csv (primary, one row per alert, 9,200 alerts)
soc_topology.csv (per-analyst registry; reserved for future
work - 25 analysts is too small to be a
useful join target beyond the analyst_tier
column already on soc_alerts)
incident_summary.csv (per-incident aggregates; reserved - only
9% of alerts link to an incident)
alert_events.csv (discrete alert event log; reserved)
Target classes (5):
auto_resolved_soar, duplicate_merged, false_positive_closed,
true_positive_escalated, true_positive_remediated
Grouping decision
-----------------
There is no natural row-level group key for CYB008:
- 25 analysts -> group-aware split would yield ~4 test analysts
- 5 SOCs -> group-aware split would yield ~1 test SOC
- 589 incidents -> only 9% of alerts have a non-null incident_id
This baseline uses STRATIFIED random splitting (like CYB001 for network
flows), which is the right choice when alerts are independent given
features. The model card documents this rationale.
Leakage audit
-------------
Three columns are structural oracles for resolution_outcome and are
DROPPED from the feature set:
1. `alert_lifecycle_phase` (4 values: auto_closed, escalated, resolved,
suppressed_duplicate): three of the four values map deterministically
to specific resolution_outcome classes. Drop.
2. `automation_resolved` (binary): exactly 1:1 with auto_resolved_soar
outcome. Drop.
3. `escalation_flag` (binary): near-1:1 with true_positive_escalated
outcome (1319 escalation flags = 1319 escalated outcomes). Drop.
With all three dropped, accuracy drops from 100% to 79% - confirming
they were structural oracles, not real predictive signal.
`soar_playbook_triggered` is a PARTIAL oracle (one-way necessary
condition: auto_resolved_soar => soar_playbook_triggered=1, but
soar_playbook_triggered=1 also yields 32% non-auto-resolve outcomes).
This is a legitimate observable - a SOAR playbook actually executing
is part of how the alert is triaged. KEPT.
`mitre_technique_id` is a perfect oracle for mitre_tactic (every T-
number belongs to one tactic by ATT&CK design) but has no relationship
to resolution_outcome. It is high-cardinality (36 values from a small
sample of a 600+-value enterprise space) and contributes nothing to
this task. Dropped for parsimony.
`detection_rule_id` has 656 unique values - too high-cardinality for
one-hot encoding. Dropped.
Identifier / non-feature columns
--------------------------------
Dropped: alert_id, incident_id (mostly null), analyst_id, soc_id,
shift_id, alert_timestamp_min, soar_playbook_id (high cardinality).
Public API
----------
build_features(alerts_path) -> (X, y, ids, meta)
transform_single(record, meta) -> np.ndarray
save_meta(meta, path) / load_meta(path)
License
-------
Ships with the public model on Hugging Face under CC-BY-NC-4.0,
matching the dataset license. See README.md.
"""
from __future__ import annotations
import json
from pathlib import Path
from typing import Any
import numpy as np
import pandas as pd
# ---------------------------------------------------------------------------
# Label space
# ---------------------------------------------------------------------------
# Ordered by triage spectrum: auto -> dup -> FP -> TP-remediate -> TP-escalate
LABEL_ORDER = [
"auto_resolved_soar",
"duplicate_merged",
"false_positive_closed",
"true_positive_remediated",
"true_positive_escalated",
]
LABEL_TO_INT = {lbl: i for i, lbl in enumerate(LABEL_ORDER)}
INT_TO_LABEL = {i: lbl for lbl, i in LABEL_TO_INT.items()}
# ---------------------------------------------------------------------------
# Identifier and target columns
# ---------------------------------------------------------------------------
ID_COLUMNS = [
"alert_id", "incident_id", "analyst_id", "soc_id", "shift_id",
"alert_timestamp_min", "soar_playbook_id",
]
TARGET_COLUMN = "resolution_outcome"
# Structural oracle columns - dropped from features.
ORACLE_COLUMNS = [
"alert_lifecycle_phase", # deterministically maps to 3 of 5 outcomes
"automation_resolved", # 1:1 with auto_resolved_soar outcome
"escalation_flag", # 1:1 with true_positive_escalated outcome
]
# High-cardinality categorical columns - dropped for tractability.
HIGH_CARDINALITY_COLUMNS = [
"mitre_technique_id", # 36 values; no relationship to outcome
"detection_rule_id", # 656 values; one-hot explosion
]
DROPPED_FROM_FEATURES = ORACLE_COLUMNS + HIGH_CARDINALITY_COLUMNS
# ---------------------------------------------------------------------------
# Per-alert numeric features
# ---------------------------------------------------------------------------
DIRECT_NUMERIC_FEATURES = [
"raw_score",
"enriched_score",
"time_in_phase_minutes",
"queue_depth_at_ingestion",
"soar_playbook_triggered", # partial oracle, kept as observable
"sla_breached_flag",
"mttd_minutes",
"mttr_minutes",
"fatigue_score_at_alert",
]
CATEGORICAL_FEATURES = [
"alert_severity", # 7 values
"alert_source", # 8 values
"mitre_tactic", # 12 values
"analyst_tier", # 3 values (alerts) / 4 (topology) -- 3 here
"siem_platform", # 8 values
]
# ---------------------------------------------------------------------------
# Engineered features
# ---------------------------------------------------------------------------
def _add_engineered_features(df: pd.DataFrame) -> pd.DataFrame:
"""
Six engineered features encoding triage-outcome hypotheses.
Each composite is a quantity a SOC analyst would compute by hand
to assess an alert's likely disposition.
"""
df = df.copy()
# 1. Enrichment lift: how much enrichment improved the raw score.
# Positive lift = enrichment increased confidence (often -> TP).
df["enrichment_lift"] = (
df["enriched_score"] - df["raw_score"]
).astype(float)
# 2. Log-scaled MTTR. MTTR is heavy-tailed (auto-resolves seconds,
# escalations hours). log1p compresses for both XGBoost and MLP.
df["log_mttr"] = np.log1p(df["mttr_minutes"].clip(lower=0)).astype(float)
# 3. Log-scaled MTTD. Same heavy-tail shape.
df["log_mttd"] = np.log1p(df["mttd_minutes"].clip(lower=0)).astype(float)
# 4. Queue pressure: queue depth times analyst fatigue. High =
# overloaded analyst, more likely to auto-resolve or escalate.
df["queue_pressure"] = (
df["queue_depth_at_ingestion"] * df["fatigue_score_at_alert"]
).astype(float)
# 5. Triage time efficiency: enrichment_score per minute in phase.
df["enrichment_per_minute"] = (
df["enriched_score"] / df["time_in_phase_minutes"].clip(lower=0.1)
).astype(float)
# 6. Is high-confidence alert: enriched score above 0.7 typically
# indicates a strong signal that warrants escalation.
df["is_high_confidence"] = (df["enriched_score"] > 0.7).astype(int)
return df
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def build_features(
alerts_path: str | Path,
) -> tuple[pd.DataFrame, pd.Series, pd.Series, dict[str, Any]]:
"""
Load soc_alerts.csv, drop target + identifiers + oracle columns,
engineer features, one-hot encode, return (X, y, ids, meta).
`ids` is a Series of alert_id values aligned with X (used for
round-tripping; not a group label since this task uses stratified
random splitting).
"""
alerts = pd.read_csv(alerts_path)
y = alerts[TARGET_COLUMN].map(LABEL_TO_INT)
if y.isna().any():
bad = alerts.loc[y.isna(), TARGET_COLUMN].unique()
raise ValueError(f"Unknown resolution_outcome values: {bad}")
y = y.astype(int)
ids = alerts["alert_id"].copy()
alerts = alerts.drop(
columns=ID_COLUMNS + [TARGET_COLUMN] + DROPPED_FROM_FEATURES,
errors="ignore",
)
alerts = _add_engineered_features(alerts)
numeric_features = (
DIRECT_NUMERIC_FEATURES
+ [
"enrichment_lift", "log_mttr", "log_mttd",
"queue_pressure", "enrichment_per_minute", "is_high_confidence",
]
)
numeric_features = [c for c in numeric_features if c in alerts.columns]
X_numeric = alerts[numeric_features].astype(float)
categorical_levels: dict[str, list[str]] = {}
blocks: list[pd.DataFrame] = []
for col in CATEGORICAL_FEATURES:
if col not in alerts.columns:
continue
levels = sorted(alerts[col].dropna().unique().tolist())
categorical_levels[col] = levels
block = pd.get_dummies(
alerts[col].astype("category").cat.set_categories(levels),
prefix=col, dummy_na=False,
).astype(int)
blocks.append(block)
X = pd.concat(
[X_numeric.reset_index(drop=True)]
+ [b.reset_index(drop=True) for b in blocks],
axis=1,
).fillna(0.0)
meta = {
"feature_names": X.columns.tolist(),
"numeric_features": numeric_features,
"categorical_levels": categorical_levels,
"label_to_int": LABEL_TO_INT,
"int_to_label": INT_TO_LABEL,
"oracle_excluded": ORACLE_COLUMNS,
"high_cardinality_excluded": HIGH_CARDINALITY_COLUMNS,
}
return X, y, ids, meta
def transform_single(
record: dict | pd.DataFrame,
meta: dict[str, Any],
) -> np.ndarray:
"""Encode a single alert record for inference."""
if isinstance(record, dict):
df = pd.DataFrame([record.copy()])
else:
df = record.copy()
df = _add_engineered_features(df)
numeric = pd.DataFrame({
col: df.get(col, pd.Series([0.0] * len(df))).astype(float).values
for col in meta["numeric_features"]
})
blocks: list[pd.DataFrame] = [numeric]
for col, levels in meta["categorical_levels"].items():
val = df.get(col, pd.Series([None] * len(df)))
block = pd.get_dummies(
val.astype("category").cat.set_categories(levels),
prefix=col, dummy_na=False,
).astype(int)
for lvl in levels:
cname = f"{col}_{lvl}"
if cname not in block.columns:
block[cname] = 0
block = block[[f"{col}_{lvl}" for lvl in levels]]
blocks.append(block)
X = pd.concat(blocks, axis=1).fillna(0.0)
X = X.reindex(columns=meta["feature_names"], fill_value=0.0)
return X.values.astype(np.float32)
def save_meta(meta: dict[str, Any], path: str | Path) -> None:
serializable = {
"feature_names": meta["feature_names"],
"numeric_features": meta["numeric_features"],
"categorical_levels": meta["categorical_levels"],
"label_to_int": meta["label_to_int"],
"int_to_label": {str(k): v for k, v in meta["int_to_label"].items()},
"oracle_excluded": meta.get("oracle_excluded", []),
"high_cardinality_excluded": meta.get("high_cardinality_excluded", []),
}
with open(path, "w") as f:
json.dump(serializable, f, indent=2)
def load_meta(path: str | Path) -> dict[str, Any]:
with open(path) as f:
meta = json.load(f)
meta["int_to_label"] = {int(k): v for k, v in meta["int_to_label"].items()}
return meta
if __name__ == "__main__":
import sys
base = Path(sys.argv[1]) if len(sys.argv) > 1 else Path("/mnt/user-data/uploads")
X, y, ids, meta = build_features(base / "soc_alerts.csv")
print(f"X shape: {X.shape}")
print(f"y shape: {y.shape}")
print(f"n_features: {len(meta['feature_names'])}")
print(f"label distribution:\n{y.map(INT_TO_LABEL).value_counts()}")
print(f"X has NaN: {X.isnull().any().any()}")