Patch XGBoost SHAP base_score bug
Browse files- chatbot/dual_inference.py +17 -5
chatbot/dual_inference.py
CHANGED
|
@@ -200,11 +200,23 @@ def run_xgboost_inference(glake_id: str, weather_seq_df: pd.DataFrame):
|
|
| 200 |
X = xgb_input[features_used]
|
| 201 |
prob = float(model.predict_proba(X)[0, 1])
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 208 |
|
| 209 |
return {
|
| 210 |
"probability": prob,
|
|
|
|
| 200 |
X = xgb_input[features_used]
|
| 201 |
prob = float(model.predict_proba(X)[0, 1])
|
| 202 |
|
| 203 |
+
try:
|
| 204 |
+
# Fix known SHAP bug with XGBoost 2.0+ where base_score is saved as a string array e.g. "[0.39]"
|
| 205 |
+
import json
|
| 206 |
+
booster = model.get_booster()
|
| 207 |
+
config = json.loads(booster.save_config())
|
| 208 |
+
base_score = config.get('learner', {}).get('learner_model_param', {}).get('base_score', '')
|
| 209 |
+
if isinstance(base_score, str) and base_score.startswith('[') and base_score.endswith(']'):
|
| 210 |
+
config['learner']['learner_model_param']['base_score'] = base_score.strip('[]')
|
| 211 |
+
booster.load_config(json.dumps(config))
|
| 212 |
+
|
| 213 |
+
explainer = shap.TreeExplainer(model)
|
| 214 |
+
shap_values = explainer.shap_values(X)
|
| 215 |
+
feature_impacts = list(zip(features_used, shap_values[0]))
|
| 216 |
+
positive_drivers = sorted([f for f in feature_impacts if f[1] > 0], key=lambda x: x[1], reverse=True)[:3]
|
| 217 |
+
except Exception as e:
|
| 218 |
+
print(f"Warning: SHAP explainability skipped due to error: {e}")
|
| 219 |
+
positive_drivers = []
|
| 220 |
|
| 221 |
return {
|
| 222 |
"probability": prob,
|