为何 SHAP 解释会随样本和模型变化

# XAI Limitations: Explanation Instability


# XAI Limitations: Explanation Instability

**Chapter 12, Section 12.5**: Model Explainability with SHAP

## Purpose
This notebook demonstrates critical limitations of model explanations
that practitioners must understand before relying on SHAP for
downstream decisions (feature pruning, risk allocation, reporting).

## Learning Objectives
- Identify explanation instability: similar predictions with different SHAP profiles
- Demonstrate the Rashomon effect across model families (GBM vs Random Forest)
- Understand when SHAP attributions reflect model-specific fitting, not data truth

## Cross-References
- **Section 12.5**: Rashomon effect, explanation multiplicity
- **Related**: `08_shap_analysis` (SHAP fundamentals), `11_conformal_gbm` (UQ)

## 1. Setup

```python
"""XAI Limitations - explanation instability across seeds and model families."""

import warnings

# lightgbm loads before anything that pulls in scikit-learn, ml4t.diagnostic included:
# the first OpenMP runtime loaded wins the process, and the wrong order segfaults on
# macOS ARM64.
import lightgbm as lgb
import matplotlib.pyplot as plt
import numpy as np
import polars as pl
from ml4t.diagnostic.metrics import cross_sectional_ic_series

# LightGBM records synthetic feature names when fitted on an array with an eval_set,
# and sklearn then warns at every predict on an array that has none to compare. One
# message, not the category: the fit and the predictions are unaffected.
warnings.filterwarnings(
    "ignore",
    message="X does not have valid feature names",
    category=UserWarning,
    module="sklearn.utils.validation",
)

import shap
from sklearn.ensemble import RandomForestRegressor

from utils.modeling import load_modeling_dataset
from utils.reproducibility import set_global_seeds
from utils.style import COLOR_CYCLER, show_with_alt
```

```python
# Above the assignment: papermill drops a parameters line whose comment contains an `=`.
MAX_SYMBOLS = 0  # the full universe
SEED = 42
```

```python
set_global_seeds(SEED)
```

## 2. Load Data

```python
mds = load_modeling_dataset("etfs", "fwd_ret_21d", max_symbols=MAX_SYMBOLS)
df = mds.dataset.to_pandas()
date_col = mds.date_col
FEATURE_COLS = mds.feature_names

# Use first walk-forward fold
split = mds.splits[0]
train_mask = (df[date_col] >= split["train_start"]) & (df[date_col] <= split["train_end"])
test_mask = (df[date_col] >= split["val_start"]) & (df[date_col] <= split["val_end"])

X_train = df.loc[train_mask, FEATURE_COLS].values
y_train = df.loc[train_mask, mds.label_col].values
X_test = df.loc[test_mask, FEATURE_COLS].values
y_test = df.loc[test_mask, mds.label_col].values

# Drop NaN labels
train_valid = np.isfinite(y_train)
test_valid = np.isfinite(y_test)
X_train, y_train = X_train[train_valid], y_train[train_valid]
X_test, y_test = X_test[test_valid], y_test[test_valid]

# Test dates and symbols for cross-sectional IC
test_entity_col = mds.entity_cols[0]
dates_test = df.loc[test_mask, date_col].values[test_valid]
symbols_test = df.loc[test_mask, test_entity_col].values[test_valid]


def cross_sectional_ic_mean(y_true, y_pred, dates, symbols):
    pred_df = pl.DataFrame({"timestamp": dates, "symbol": symbols, "prediction": y_pred})
    ret_df = pl.DataFrame({"timestamp": dates, "symbol": symbols, "forward_return": y_true})
    ic_per_date = cross_sectional_ic_series(
        pred_df,
        ret_df,
        pred_col="prediction",
        ret_col="forward_return",
        date_col="timestamp",
        entity_col="symbol",
    )
    ic_clean = ic_per_date.drop_nulls("ic")
    return float(ic_clean["ic"].mean()) if ic_clean.height else float("nan")


print(f"ETFs: {len(X_train):,} train / {len(X_test):,} test ({len(FEATURE_COLS)} features)")
```

## 3. Demonstration 1: Similar Predictions, Different Explanations

Two samples with nearly identical predictions can have entirely
different SHAP profiles: the model arrived at the same answer
through different reasoning paths.

```python
model = lgb.LGBMRegressor(n_estimators=100, max_depth=4, random_state=SEED, verbose=-1)
model.fit(X_train, y_train)

explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)
predictions = model.predict(X_test)
```

```python
# Find pairs with similar predictions but different SHAP
n_samples = min(500, len(predictions))
prediction_diffs = np.abs(predictions[:n_samples, None] - predictions[None, :n_samples])
shap_diffs = np.sqrt(
    ((shap_values[:n_samples, None, :] - shap_values[None, :n_samples, :]) ** 2).sum(axis=2)
)

# Use generous thresholds: prediction diff in bottom 25%, SHAP diff in top 75%
pred_threshold = np.percentile(prediction_diffs[prediction_diffs > 0], 25)
shap_threshold = np.percentile(shap_diffs[shap_diffs > 0], 75)

mask = (prediction_diffs < pred_threshold) & (shap_diffs > shap_threshold) & (prediction_diffs > 0)
pairs = np.argwhere(mask)
```

```python
print(f"Searched {n_samples} samples: {len(pairs)} instability pairs found")
print(f"  Prediction threshold: {pred_threshold:.6f}")
print(f"  SHAP distance threshold: {shap_threshold:.4f}")

if len(pairs) > 0:
    i, j = pairs[0]
    instability_df = pl.DataFrame(
        {
            "feature": FEATURE_COLS,
            f"SHAP (sample {i})": shap_values[i].round(4),
            f"SHAP (sample {j})": shap_values[j].round(4),
            "abs_diff": np.abs(shap_values[i] - shap_values[j]).round(4),
        }
    ).sort("abs_diff", descending=True)

    print(f"\nExample: samples {i} and {j}")
    print(f"  Prediction diff: {abs(predictions[i] - predictions[j]):.6f}")
    print(f"  SHAP Euclidean distance: {shap_diffs[i, j]:.4f}")
else:
    instability_df = pl.DataFrame(
        schema={
            "feature": str,
            "SHAP (sample a)": float,
            "SHAP (sample b)": float,
            "abs_diff": float,
        }
    )
    print("\nNo pairs met both thresholds; predictions and SHAP are tightly coupled")
    print("in this dataset. This is informative: it means the model's explanation")
    print("space is relatively stable for the ETF feature set.")
instability_df.head(10)
```

**Interpretation**: SHAP explanations are not unique. The same prediction
can be reached through different feature contribution paths. When instability
pairs are found, the top-three SHAP contributors can differ completely even
when predictions agree to within fractions of a percent. For production
deployment, report confidence intervals on feature importance rather than
point estimates.

## 4. Demonstration 2: Rashomon Effect Across Model Families

Models with similar predictive performance can attribute predictions
to different features. We compare three LightGBM seeds plus a
Random Forest to demonstrate cross-architecture disagreement.

```python
# Stochastic training, so different seeds give different models.
_lgb_kw = dict(
    n_estimators=200,
    max_depth=6,
    learning_rate=0.05,
    subsample=0.8,
    colsample_bytree=0.8,
    verbose=-1,
)
model_configs = [
    ("LightGBM (seed=42)", lgb.LGBMRegressor(**_lgb_kw, random_state=42)),
    ("LightGBM (seed=123)", lgb.LGBMRegressor(**_lgb_kw, random_state=123)),
    ("LightGBM (seed=456)", lgb.LGBMRegressor(**_lgb_kw, random_state=456)),
    (
        "RandomForest",
        RandomForestRegressor(n_estimators=200, max_depth=6, random_state=SEED, n_jobs=-1),
    ),
]
```

```python
model_results = []
importance_arrays = []

for name, m in model_configs:
    m.fit(X_train, y_train)
    pred = m.predict(X_test)
    ic = cross_sectional_ic_mean(y_test, pred, dates_test, symbols_test)

    exp = shap.TreeExplainer(m)
    sv = exp.shap_values(X_test)
    mean_abs = np.abs(sv).mean(axis=0)

    model_results.append({"model": name, "ic": round(ic, 4)})
    importance_arrays.append(mean_abs)
```

```python
rashomon_df = pl.DataFrame(model_results)
rashomon_df
```

Read the table by the spread rather than by any single value. The three
LightGBM seeds differ only in their random seed, and they span a wider range
than separates any of them from the Random Forest, which lands inside that
span. Changing the seed therefore moves validation IC by more than changing
the model family does here. The Rashomon point holds in either direction:
similar predictive performance, different feature attributions.

```python
# Show top-5 features per model
for i, (name, _) in enumerate(model_configs):
    order = np.argsort(importance_arrays[i])[::-1][:5]
    top5 = [(FEATURE_COLS[idx], round(importance_arrays[i][idx], 4)) for idx in order]
    print(f"{name}: {', '.join(f'{f} ({v})' for f, v in top5)}")
```

```python
# Grouped bar chart comparing SHAP importance profiles
fig, ax = plt.subplots(figsize=(12, 5))
n_feat = min(10, len(FEATURE_COLS))
top_feats_idx = np.argsort(importance_arrays[0])[::-1][:n_feat]
feat_names = [FEATURE_COLS[idx] for idx in top_feats_idx]

x = np.arange(n_feat)
width = 0.2
colors = COLOR_CYCLER[:4]  # four distinct categorical hues (blue, amber, copper, green)

for i, (name, _) in enumerate(model_configs):
    vals = [importance_arrays[i][idx] for idx in top_feats_idx]
    ax.bar(x + i * width, vals, width, label=name, color=colors[i])

ax.set_xticks(x + width * 1.5)
ax.set_xticklabels(feat_names, rotation=45, ha="right", fontsize=9)
ax.set_ylabel("Mean |SHAP|")
ax.set_title("Mean absolute SHAP by feature, across seeds and model families")
ax.legend(fontsize=8)
show_with_alt(
    fig,
    "Grouped bars of mean absolute SHAP value, one group per feature and one bar per "
    "model, comparing three LightGBM seeds against a Random Forest.",
)
```

**Interpretation**: Features that rank highly across all four models
(both GBM and Random Forest) are more likely to reflect genuine data
structure. Features that only one family highlights may reflect
architecture-specific fitting patterns. When SHAP attributions
drive downstream decisions, validate across model specifications.

## 5. Stakeholder Guidance

| Audience | Recommended Explanation | Key Caveat |
|----------|------------------------|------------|
| Risk Manager | SHAP summary plot | Correlation $\neq$ causation |
| Trader | Feature importance rank | May vary with retraining |
| Regulator | Model documentation | Include confidence intervals |
| Quant Researcher | Full SHAP + interactions | Check stability across seeds |

## 6. Key Takeaways

1. **Explanation instability**: Samples with similar predictions can have
   different top SHAP contributors: the same output is reachable through
   different feature-contribution paths.

2. **Rashomon effect**: Models with different architectures or random seeds
   attribute predictions to different features, so SHAP explanations reflect
   model-specific fitting patterns, not ground truth about the data.

3. **Best practices**: Report confidence intervals on feature importance,
   check stability across random seeds and model architectures, and never
   claim SHAP proves causation.

**Next**: See `08_shap_analysis` for SHAP fundamentals and drift detection,
or Section 12.5 in the chapter text for the theoretical framework.
![notebook output](figures/p1_1.png)