Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions sentinelml/explainability/root_cause.py
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,14 @@ def rank_contributing_features(
or values.shape[1] != len(self.feature_names)
or not np.isfinite(values).all()):
raise ValueError("SHAP attributions must be finite nonempty observation-by-feature matrices")
baseline_mean_abs = np.mean(np.abs(baseline_shap_values), axis=0)
degraded_mean_abs = np.mean(np.abs(degraded_shap_values), axis=0)
def mean_abs(values):
absolute = np.abs(values)
scale = np.max(absolute, axis=0)
normalized = np.divide(absolute, scale, out=np.zeros_like(absolute), where=scale != 0)
return np.mean(normalized, axis=0) * scale

baseline_mean_abs = mean_abs(baseline_shap_values)
degraded_mean_abs = mean_abs(degraded_shap_values)

shift_magnitudes = np.abs(degraded_mean_abs - baseline_mean_abs)
top_indices = np.argsort(shift_magnitudes)[::-1][:top_k]
Expand Down
56 changes: 23 additions & 33 deletions src/root_cause.py
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,22 @@
}



def _mean_absolute_attributions(values):
absolute = np.abs(values)
scale = np.max(absolute, axis=0)
normalized = np.divide(absolute, scale, out=np.zeros_like(absolute), where=scale != 0)
return np.mean(normalized, axis=0) * scale


def _importance_percentages(values):
scale = np.max(values)
if scale == 0:
return np.zeros_like(values)
normalized = values / scale
return normalized / normalized.sum() * 100


def analyze_root_causes():

print("\nSENTINELML — ROOT CAUSE ANALYSIS")
Expand Down Expand Up @@ -87,9 +103,7 @@ def analyze_root_causes():
# 2. Global feature importance
# --------------------------------

mean_abs_shap = np.abs(
shap_values
).mean(axis=0)
mean_abs_shap = _mean_absolute_attributions(shap_values)

importance = pd.DataFrame(
{
Expand All @@ -103,19 +117,10 @@ def analyze_root_causes():
ascending=False,
)

total_importance = (
importance["mean_abs_shap"].sum()
importance["importance_percent"] = _importance_percentages(
importance["mean_abs_shap"].to_numpy()
)

if total_importance > 0:
importance["importance_percent"] = (
importance["mean_abs_shap"]
/ total_importance
* 100
)
else:
importance["importance_percent"] = 0

# --------------------------------
# 3. Analyze HIGH-risk predictions
# --------------------------------
Expand All @@ -136,26 +141,11 @@ def analyze_root_causes():
shap_values[high_risk_mask]
)

high_risk_importance = (
high_risk_shap.mean(axis=0)
)

total_high = (
high_risk_importance.sum()
)

for feature, value in zip(
features,
high_risk_importance,
):

percent = (
float(value / total_high * 100)
if total_high > 0
else 0.0
)
high_risk_importance = _mean_absolute_attributions(high_risk_shap)
percentages = _importance_percentages(high_risk_importance)

high_risk_causes[feature] = percent
for feature, percent in zip(features, percentages):
high_risk_causes[feature] = float(percent)

high_risk_causes = dict(
sorted(
Expand Down
16 changes: 16 additions & 0 deletions tests/test_explainability_extreme_values.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@
import numpy as np
import pytest

from sentinelml.explainability.root_cause import RootCauseExplainer


def test_attribution_ranking_handles_extreme_finite_values():
explainer = RootCauseExplainer(["a", "b"])
baseline = np.array([[1e308, 1e308], [1e308, 1e308]])
degraded = np.array([[5e307, 1e308], [5e307, 1e308]])
with np.errstate(over="raise", invalid="raise"):
result = explainer.rank_contributing_features(baseline, degraded, top_k=2)
assert result[0]["feature"] == "a"
assert result[0]["shift_magnitude"] == pytest.approx(5e307)
assert result[1]["shift_magnitude"] == 0.0
assert all(np.isfinite(item["baseline_importance"]) for item in result)
40 changes: 40 additions & 0 deletions tests/test_root_cause_extreme_values.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,40 @@
import json

import numpy as np
import pandas as pd
import pytest

from src import root_cause


@pytest.mark.parametrize(
"values, expected",
[
([[1e308, 1e308], [1e308, 1e308]], {"a": 50, "b": 50}),
([[0., 0.], [0., 0.]], {"a": 0, "b": 0}),
([[1., 3.], [1., 3.]], {"a": 25, "b": 75}),
],
)
def test_finite_shap_produces_normalized_feature_importance(tmp_path, monkeypatch, values, expected):
production = tmp_path / "production.csv"
pd.DataFrame({"a": [1., 2.], "b": [3., 4.], "prediction_risk": ["HIGH", "HIGH"]}).to_csv(production, index=False)
monkeypatch.setattr(root_cause, "PRODUCTION_PATH", production)
monkeypatch.setattr(root_cause, "OUTPUT_PATH", tmp_path / "causes.json")

class FakeModel:
feature_names_in_ = ["a", "b"]

class FakeExplainer:
def __init__(self, model):
pass

def shap_values(self, frame):
return np.asarray(values)

monkeypatch.setattr(root_cause.joblib, "load", lambda _: FakeModel())
monkeypatch.setattr(root_cause.shap, "TreeExplainer", FakeExplainer)
with np.errstate(over="raise", invalid="raise"):
root_cause.analyze_root_causes()
report = json.loads((tmp_path / "causes.json").read_text())
assert report["global_feature_importance"] == pytest.approx(expected)
assert report["high_risk_root_causes"] == pytest.approx(expected)
Loading