Feature importance in regression in ML Python - Deep Dive

Overview - Feature importance in regression

What is it?

Feature importance in regression tells us how much each input variable (feature) affects the prediction of a regression model. It helps us understand which features are most useful for predicting the target value. This is done by measuring the impact of each feature on the model's accuracy or output. Knowing feature importance helps us trust and improve our models.

Why it matters

Without knowing feature importance, we treat all input features as equally valuable, which can hide the true drivers of predictions. This can lead to models that are harder to interpret, less efficient, or even biased. Feature importance helps us focus on the most meaningful data, improving model performance and making decisions clearer and more reliable in real life, like predicting house prices or sales.

Where it fits

Before learning feature importance, you should understand basic regression concepts and how models make predictions. After this, you can explore feature selection, model interpretation techniques, and advanced explainability methods like SHAP or LIME.

Mental Model

Core Idea

Feature importance measures how much each input feature influences the prediction outcome in a regression model.

Think of it like...

Imagine baking a cake with many ingredients; feature importance is like knowing which ingredients affect the cake's taste the most.

Regression Model
  ├─ Feature 1 (Importance: High)
  ├─ Feature 2 (Importance: Medium)
  ├─ Feature 3 (Importance: Low)
  └─ Feature 4 (Importance: None)

Importance shows how much each feature 'pulls' the prediction.

Build-Up - 7 Steps

1

Foundation

Understanding regression basics

Concept: Learn what regression models do and how they predict continuous values.

Regression models predict a number based on input features. For example, predicting house prices from size and location. The model learns a formula that connects inputs to outputs.

Result

You understand that regression outputs numbers and depends on input features.

Knowing how regression works is essential before measuring which features matter most.

2

Foundation

What are features in regression?

Concept: Features are the input variables used to make predictions in regression.

Features can be things like age, size, or temperature. Each feature provides information that helps the model guess the target value.

Result

You can identify features and understand their role in prediction.

Recognizing features as inputs clarifies why some might be more important than others.

3

Intermediate

Measuring feature importance by coefficients

🤔Before reading on: do you think larger coefficients always mean more important features? Commit to yes or no.

Concept: In linear regression, feature importance can be estimated by the size of coefficients assigned to each feature.

Linear regression assigns a number (coefficient) to each feature showing how much the prediction changes if that feature changes by one unit. Larger absolute coefficients usually mean more influence.

Result

You can interpret coefficients as a simple measure of feature importance.

Understanding coefficients connects the math of regression to the idea of importance.

4

Intermediate

Using permutation importance for any model

🤔Before reading on: do you think shuffling a feature's values affects model error if the feature is unimportant? Commit to yes or no.

Concept: Permutation importance measures how much model error increases when a feature's values are randomly shuffled.

By shuffling one feature's values, we break its link to the target. If error grows a lot, that feature was important. If error stays the same, the feature was not important.

Result

You can estimate feature importance for any regression model, even complex ones.

Knowing permutation importance lets you measure importance without relying on model internals.

5

Intermediate

Feature importance with tree-based models

Concept: Tree models calculate importance by how much each feature reduces prediction error when splitting data.

Decision trees split data based on features to reduce error. Features that split data better get higher importance scores. These scores can be averaged over many trees in ensembles like random forests.

Result

You understand how tree models provide built-in feature importance.

Seeing importance as error reduction links model structure to feature influence.

6

Advanced

Limitations of feature importance methods

🤔Before reading on: do you think correlated features always get equal importance? Commit to yes or no.

Concept: Feature importance can be misleading when features are correlated or when models are complex.

If two features provide similar information, importance may be split or assigned unevenly. Some methods may overstate or understate importance. Understanding these limits helps avoid wrong conclusions.

Result

You recognize when feature importance might not tell the full story.

Knowing limitations prevents misuse and builds trust in model explanations.

7

Expert

Advanced explainability with SHAP values

🤔Before reading on: do you think SHAP values assign importance fairly even with correlated features? Commit to yes or no.

Concept: SHAP values provide a unified way to fairly distribute prediction contributions among features, even when correlated.

SHAP uses game theory to assign each feature a contribution value for each prediction. It considers all feature combinations, giving consistent and fair importance scores.

Result

You can explain individual predictions and global importance with a solid theoretical foundation.

Understanding SHAP reveals how to overcome common pitfalls in feature importance interpretation.

Under the Hood

Feature importance methods work by measuring how changes in a feature affect the model's predictions or error. For linear models, coefficients directly scale features. For permutation importance, shuffling breaks the feature-target link, increasing error if important. Tree-based methods track error reduction at splits. SHAP values compute contributions by averaging over all feature subsets, ensuring fair attribution.

Why designed this way?

These methods were designed to provide interpretable insights into complex models. Linear coefficients are simple but limited to linear relationships. Permutation importance is model-agnostic but can be biased by feature correlation. Tree-based importance leverages model structure for efficiency. SHAP was created to unify and improve fairness in attribution, addressing earlier methods' shortcomings.

Feature Importance Methods
┌─────────────────────────────┐
│ Linear Regression           │
│  Coefficients → Importance │
├─────────────────────────────┤
│ Permutation Importance      │
│  Shuffle feature → Measure  │
│  error increase             │
├─────────────────────────────┤
│ Tree-based Importance       │
│  Sum error reduction at     │
│  splits                    │
├─────────────────────────────┤
│ SHAP Values                 │
│  Average contributions over │
│  all feature subsets        │
└─────────────────────────────┘

Myth Busters - 4 Common Misconceptions

Quick: does a large coefficient always mean a feature is the most important? Commit to yes or no.

Common Belief: A large coefficient in linear regression means the feature is the most important.

Reality: Coefficient size alone can be misleading because features with different scales or correlations affect coefficients. Importance should consider feature scale and context.

Why it matters: Misinterpreting coefficients can lead to wrong conclusions about which features truly influence predictions.

Quick: does permutation importance always give unbiased importance scores? Commit to yes or no.

Common Belief: Permutation importance always accurately reflects feature importance regardless of feature relationships.

Reality: Permutation importance can be biased when features are correlated, as shuffling one breaks relationships and may overstate or understate importance.

Why it matters: Ignoring this can cause wrong feature selection or misunderstanding of model behavior.

Quick: do tree-based feature importance scores always reflect true feature influence? Commit to yes or no.

Common Belief: Tree-based importance scores perfectly show how important each feature is for prediction.

Reality: Tree importance can be biased toward features with more split options or higher cardinality, not always reflecting true importance.

Why it matters: Relying blindly on these scores can mislead feature engineering and interpretation.

Quick: can feature importance explain individual predictions? Commit to yes or no.

Common Belief: Feature importance always explains why a single prediction was made.

Reality: Most feature importance methods explain global influence, not individual predictions. Methods like SHAP are needed for local explanations.

Why it matters: Confusing global and local explanations can cause misunderstanding of model decisions.

Expert Zone

1

Feature importance can vary depending on the data sample and model randomness, so stability checks are crucial.

2

Correlated features can share importance, making it hard to assign credit uniquely; advanced methods like SHAP address this.

3

Feature importance does not imply causation; a feature can be important without causing the target.

When NOT to use

Feature importance is less reliable when features are highly correlated or when the model is unstable. In such cases, use dimensionality reduction, causal inference methods, or robust explainability tools like SHAP or LIME.

Production Patterns

In production, feature importance guides feature selection to reduce model complexity and improve speed. It also helps monitor model drift by tracking changes in feature influence over time. Explainability reports using SHAP are often integrated into dashboards for stakeholder trust.

Connections

Causal inference

Builds-on

Understanding feature importance helps identify which variables influence predictions, a step toward discovering causal relationships.

Dimensionality reduction

Complementary

Feature importance guides which features to keep or remove before applying dimensionality reduction techniques like PCA.

Game theory

Builds-on

SHAP values use game theory concepts to fairly distribute prediction contributions among features, connecting ML explainability to economic theory.

Common Pitfalls

#1 Interpreting raw coefficients as absolute importance without scaling features.

Wrong approach:

model.coef_ # directly used as importance without considering feature scale

Correct approach:

from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
model.fit(X_scaled, y)
importance = model.coef_

Root cause:

Ignoring that features with larger scales naturally have larger coefficients, misleading importance interpretation.

#2 Using permutation importance on correlated features without caution.

Wrong approach:

from sklearn.inspection import permutation_importance
result = permutation_importance(model, X, y)
print(result.importances_mean)

Correct approach:

# Check feature correlations first
import numpy as np
corr = np.corrcoef(X.T)
# Interpret permutation importance carefully or use SHAP for correlated features

Root cause:

Not accounting for feature correlation causes biased importance scores.

#3 Assuming feature importance explains individual predictions.

Wrong approach:

print(model.feature_importances_) # used to explain a single prediction

Correct approach:

import shap
explainer = shap.Explainer(model, X)
shap_values = explainer(X)
shap.plots.waterfall(shap_values[0]) # explains one prediction

Root cause:

Confusing global importance with local explanation leads to wrong interpretation of single predictions.

Key Takeaways

Feature importance reveals how much each input feature influences regression model predictions.

Different methods exist: coefficients for linear models, permutation importance for any model, and tree-based importance for decision trees.

Feature importance can be misleading when features are correlated or unscaled; advanced methods like SHAP provide fairer explanations.

Understanding feature importance helps improve model trust, simplify models, and focus on meaningful data.

Feature importance explains global model behavior but usually not individual predictions without specialized tools.