Business Analytics 101: Ethics and Privacy in Analytics Explainability LIME SHAP Interpretable Models

What This Is

Explainability in business analytics refers to the practice of making complex models and algorithms more understandable and transparent. This is crucial in business analytics as it helps stakeholders trust and interpret the results, making informed decisions. For instance, a retail company wants to predict sales using a machine learning model. By using explainability techniques, they can understand which features (e.g., seasonality, price, advertising) contribute most to the predictions, enabling them to make data-driven decisions.

Key Formulas & Metrics

• LIME (Local Interpretable Model-agnostic Explanations) Score = Σ(wᵢ * xᵢ) – weighted sum of input features that contribute to the prediction.
• wᵢ: weight of feature i
• xᵢ: value of feature i
• Σ: sum over all features
• SHAP (SHapley Additive exPlanations) Value = φᵢ = Σ(Δφᵢ / |S|) – contribution of feature i to the prediction.
• φᵢ: SHAP value of feature i
• Δφᵢ: change in prediction when feature i is added or removed
• |S|: number of features in the model
• Interpretable Model Coefficient = βᵢ = (Σ(xᵢ * yᵢ) - (Σxᵢ) * (Σyᵢ) / n) / (Σxᵢ² - (Σxᵢ)² / n) – coefficient of feature i in a linear regression model.
• βᵢ: coefficient of feature i
• xᵢ: value of feature i
• yᵢ: value of target variable
• n: number of observations
• Mean Absolute Error (MAE) = (1/n) Σ|yᵢ − ŷᵢ| – average absolute forecast error.
• yᵢ: actual value
• ŷᵢ: predicted value
• n: number of observations
• R² = 1 − (SS_res / SS_tot) – proportion of variance explained by the regression model.
• SS_res: sum of squared residuals
• SS_tot: total sum of squares
• Feature Importance = Σ(|βᵢ|) – sum of absolute coefficients of features in a linear regression model.
• βᵢ: coefficient of feature i
• |βᵢ|: absolute value of coefficient i

Step-by-Step Procedure

1. Choose a model: Select a suitable model for the business problem, such as linear regression or decision trees.
2. Collect data: Gather relevant data for the model, including input features and target variable.
3. Train the model: Train the model using the collected data.
4. Apply explainability techniques: Use techniques like LIME, SHAP, or interpretable models to understand the model's predictions.
5. Interpret results: Analyze the results of the explainability techniques to identify key features contributing to the predictions.
6. Communicate findings: Present the findings to stakeholders, highlighting the insights gained from the explainability techniques.

Common Mistakes

• Mistake: Confusing correlation with causation.
• Correction: Establish causality by using techniques like regression analysis or experimentation.
• Mistake: Misinterpreting p-values.
• Correction: Understand that p-values indicate the probability of observing the data (or more extreme) if the null hypothesis is true, not the probability that the null hypothesis is true.
• Mistake: Using the wrong error metric for a business problem.
• Correction: Choose an error metric that aligns with the business objective, such as mean absolute error for forecasting or accuracy for classification.

Software / Tool Tips

• Python with scikit-learn: Use the LIME and SHAP libraries to apply explainability techniques.
• R: Utilize the lime and shap packages for explainability.
• Tableau: Use the Interpret feature to understand the contributions of input features to the predictions.

Quick Practice Problem

Scenario: A company wants to predict sales using a linear regression model. The model has two features: price and advertising. The coefficient of price is 0.5, and the coefficient of advertising is 1.2. What does this mean?

Answer: The price feature contributes 0.5 units to the prediction for every dollar increase, while the advertising feature contributes 1.2 units for every dollar increase.

Last-Minute Cram Sheet

1. LIME scores represent the weighted sum of input features contributing to the prediction.
2. SHAP values indicate the contribution of each feature to the prediction.
3. Interpretable model coefficients represent the contribution of each feature to the prediction.
4. MAE measures the average absolute forecast error.
5. R² represents the proportion of variance explained by the regression model.
6. Feature importance is the sum of absolute coefficients of features in a linear regression model.
7. ⚠️ p-value is NOT the probability that H₀ is true – it's the probability of observing the data (or more extreme) if H₀ is true.
8. ⚠️ Correlation does not imply causation.
9. ⚠️ Use the right error metric for the business problem.
10. Interpretable models provide transparent and understandable predictions.