Feature Explanation¶

Feature explanation helps us understand how individual features impact the output — critical for feature selection, bias detection, and explaining decisions.

Information Value (IV) Plots¶

IV Plots combine Weight of Evidence (WoE) and Information Value (IV) to quantify a feature's predictive power.

Weight of Evidence (WoE)¶

\[WoE_i = \ln\left(\frac{\text{Distribution of Events}_i}{\text{Distribution of Non-Events}_i}\right)\]

Where events = favourable outcomes, non-events = unfavourable outcomes, and \(i\) represents a bin of the feature.

WoE Interpretation

WoE > 0: Bin has more events than expected → favourable
WoE < 0: Bin has fewer events than expected → unfavourable
WoE = 0: No predictive power in this bin

Information Value (IV)¶

\[IV = \sum_{i=1}^{n} (D_{events_i} - D_{nonevents_i}) \times WoE_i\]

IV Range	Predictive Power
< 0.02	Not useful
0.02 – 0.1	Weak
0.1 – 0.3	Medium
0.3 – 0.5	Strong
> 0.5	Suspicious (too good — check for overfitting)

import numpy as np
import pandas as pd

def compute_woe_iv(df, feature, target, bins=10):
    """Compute WoE and IV for a feature."""
    # Bin continuous features
    if df[feature].nunique() > bins:
        df['bin'] = pd.qcut(df[feature], q=bins, duplicates='drop')
    else:
        df['bin'] = df[feature]

    grouped = df.groupby('bin')[target].agg(['sum', 'count'])
    grouped.columns = ['events', 'total']
    grouped['non_events'] = grouped['total'] - grouped['events']

    # Distribution
    total_events = grouped['events'].sum()
    total_non_events = grouped['non_events'].sum()

    grouped['dist_events'] = grouped['events'] / total_events
    grouped['dist_non_events'] = grouped['non_events'] / total_non_events

    # WoE and IV
    grouped['woe'] = np.log(
        grouped['dist_events'].clip(0.0001) / 
        grouped['dist_non_events'].clip(0.0001)
    )
    grouped['iv'] = (grouped['dist_events'] - grouped['dist_non_events']) * grouped['woe']

    iv_total = grouped['iv'].sum()
    df.drop('bin', axis=1, inplace=True)

    return grouped, iv_total

Partial Dependence Plots (PDP)¶

PDPs show the marginal effect of a feature on the predicted outcome, averaging over all other features.

\[\hat{f}_{PDP}(x_s) = \frac{1}{n}\sum_{i=1}^{n} \hat{f}(x_s, x_{c}^{(i)})\]

Where \(x_s\) is the feature of interest and \(x_c\) are all other features.

from sklearn.inspection import PartialDependenceDisplay
import matplotlib.pyplot as plt

# After training a model
PartialDependenceDisplay.from_estimator(
    model, X_test, features=[0, 1, 2],
    kind='average', grid_resolution=50
)
plt.tight_layout()
plt.show()

SHAP Values¶

SHAP (SHapley Additive exPlanations) assigns each feature an importance value based on game theory — how much does each feature contribute to moving the prediction from the baseline?

\[\phi_i = \sum_{S \subseteq F \setminus \{i\}} \frac{|S|!(|F|-|S|-1)!}{|F|!} [f(S \cup \{i\}) - f(S)]\]

import shap

# Train your model first
explainer = shap.Explainer(model, X_train)
shap_values = explainer(X_test)

# Summary plot
shap.summary_plot(shap_values, X_test)

# Individual prediction
shap.waterfall_plot(shap_values[0])

SHAP for Fairness

Compare SHAP values across protected groups. If a protected feature (or its proxy) has high SHAP importance, your model may be learning bias.

Next: Model Explanation →