File size: 5,355 Bytes
72fdabd | 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 | import logging
import shap
import numpy as np
import pandas as pd
from datetime import datetime
from models.artifacts_loader import get_artifacts
from schemas.explain import ExplanationResponse, SHAPFeature, FeatureImportanceResponse
from utils.plot_utils import (
generate_shap_summary_plot_base64,
plot_feature_importance_heatmap,
plotly_to_base64,
)
logger = logging.getLogger(__name__)
class ExplainerService:
@staticmethod
def explain_shap(
item: dict, top_n: int = 10, plot: bool = False
) -> ExplanationResponse:
artifacts = get_artifacts()
if not artifacts.model:
raise RuntimeError("Model not loaded")
try:
# Prepare Input
model = artifacts.model
df = model._to_dataframe([item])
df_norm = model._normalize_boq_schema(df)
# The background data for SHAP is expected to be transformed data.
# Explaining the inner classifier using transformed input.
preprocessor = model.preprocessor
calibrator = model.calibrator
X_proc = preprocessor.transform(df_norm)
# Predict to find the target class index
try:
pred_label = model.predict(df)[0]
target_class_index = list(calibrator.classes_).index(pred_label)
except Exception:
pred_label = "Unknown"
target_class_index = 0
# 2. Compute SHAP
# Using KernelExplainer because CalibratedClassifierCV doesn't support TreeExplainer natively easily
if artifacts.shap_background is None:
raise RuntimeError(
"SHAP background data is missing but required for KernelExplainer"
)
explainer = shap.KernelExplainer(
calibrator.predict_proba, artifacts.shap_background
)
shap_values_all = explainer.shap_values(X_proc, nsamples=100)
# Robustly extract values for the target class
if isinstance(shap_values_all, list):
if len(shap_values_all) > target_class_index:
shap_vals_target = shap_values_all[target_class_index]
else:
shap_vals_target = shap_values_all[0]
elif isinstance(shap_values_all, np.ndarray):
if len(shap_values_all.shape) == 3 and shap_values_all.shape[2] > target_class_index:
shap_vals_target = shap_values_all[:, :, target_class_index]
else:
shap_vals_target = shap_values_all
else:
shap_vals_target = shap_values_all
if len(shap_vals_target.shape) == 2:
sv = shap_vals_target[0]
else:
sv = shap_vals_target
# 3. Features Names
try:
if hasattr(preprocessor, "get_feature_names_out"):
feature_names = preprocessor.get_feature_names_out()
else:
feature_names = preprocessor.pipeline.get_feature_names_out()
except Exception:
feature_names = [f"fet_{i}" for i in range(len(sv))]
# 4. Top Features
abs_importance = np.abs(sv)
top_indices = np.argsort(abs_importance)[-top_n:][::-1]
top_features = [
SHAPFeature(feature=str(feature_names[i]), value=float(sv[i]))
for i in top_indices
]
# 5. Plot
plot_b64 = None
if plot:
plot_b64 = generate_shap_summary_plot_base64(
shap_vals_target, X_proc, feature_names=feature_names, target_class=pred_label
)
return ExplanationResponse(
top_features=top_features,
shap_plot_base64=plot_b64,
metadata={
"method": "SHAP KernelExplainer",
"timestamp": datetime.utcnow().isoformat(),
},
)
except Exception as e:
logger.exception("Error during SHAP explanation")
raise RuntimeError(f"Explanation failed: {e}")
@staticmethod
def explain_feature_importance(
top_n: int = 30, skip_top: int = 0
) -> FeatureImportanceResponse:
artifacts = get_artifacts()
if not artifacts.model:
raise RuntimeError("Model not loaded")
try:
fig, df_plot = plot_feature_importance_heatmap(
model=artifacts.model,
top_n=top_n,
skip_top=skip_top,
title="Global Feature Importance",
)
plot_base64 = plotly_to_base64(fig)
return FeatureImportanceResponse(
plot_base64=plot_base64,
metadata={
"method": "Averaged Calibrated Feature Importances",
"timestamp": datetime.utcnow().isoformat(),
"top_n": top_n,
"skip_top": skip_top,
},
)
except Exception as e:
logger.exception("Error during feature importance generation")
raise RuntimeError(f"Feature importance generation failed: {e}")
|