Binary outcome with features A and B, corr(A,B)=0.98, plus other weakly correlated features. You fit (i) logistic regression and (ii) a gradient-boosted tree. (1) Compute/estimate VIF for A and B and interpret thresholds that indicate problematic multicollinearity. (2) Explain how SHAP values behave when A and B are near-duplicates (interventional vs conditional SHAP), and why attributions may be unstable or split unpredictably across A and B. (3) Propose a defensible interpretation workflow: feature clustering/grouped SHAP, permutation importance conditional on the other, and re-fitting after removing one of the pair; describe the diagnostics you would expect to see. (4) Recommend modeling changes (e.g., elastic net for GLM or feature grouping/penalization for trees) and how you would validate that interpretability and predictive performance are both acceptable.

This question evaluates understanding of multicollinearity diagnostics (VIF), feature attribution methods (SHAP), and the interpretability and validation implications of near-duplicate predictors in binary classification models.

How do I approach Machine Learning interview questions?

Machine Learning questions require understanding of core concepts and practice. PracHub provides solutions with explanations to help you master machine learning interviews.

What difficulty level is this interview question?

This is a hard difficulty Machine Learning question, commonly asked during Technical Screen rounds at TikTok.

What role is this question designed for?

This question is commonly asked for Data Scientist candidates at TikTok during technical interviews.

Explain SHAP vs VIF under collinearity | TikTok Interview Question

High Collinearity in Binary Classification: VIF, SHAP, and Interpretation Strategy

You are modeling a binary outcome Y. Two numeric features A and B are highly correlated: corr(A, B) = 0.98. Other features exist but are only weakly correlated with A and B. You fit:

(i) a logistic regression (GLM), and
(ii) a gradient-boosted tree model (GBDT).

Answer the following:

Variance Inflation Factor (VIF)
- Compute/estimate the VIF for A and B given corr(A, B) = 0.98 (assume other features add little to R²).
- Interpret typical VIF/tolerance thresholds that indicate problematic multicollinearity and what that means for logistic regression coefficients.
SHAP with near-duplicate features
- Explain how SHAP values behave when A and B are near-duplicates under interventional vs conditional SHAP formulations.
- Why can attributions be unstable or split unpredictably across A and B (for both GLM and GBDT)?
Defensible interpretation workflow
- Propose a workflow to interpret such models: e.g., feature clustering and grouped SHAP, permutation importance conditional on the other feature, and re-fitting after removing one of the pair.
- Describe the diagnostics you expect to see if A and B are redundant.
Modeling recommendations and validation
- Recommend modeling changes (e.g., elastic net for GLM; feature grouping or regularization choices for trees) to handle collinearity.
- Describe how you would validate that both interpretability and predictive performance remain acceptable.

High Collinearity in Binary Classification: VIF, SHAP, and Interpretation Strategy

You are modeling a binary outcome Y. Two numeric features A and B are highly correlated: corr(A, B) = 0.98. Other features exist but are only weakly correlated with A and B. You fit:

(i) a logistic regression (GLM), and

(ii) a gradient-boosted tree model (GBDT).

Answer the following:

Variance Inflation Factor (VIF)

Compute/estimate the VIF for A and B given corr(A, B) = 0.98 (assume other features add little to R²).
Interpret typical VIF/tolerance thresholds that indicate problematic multicollinearity and what that means for logistic regression coefficients.

SHAP with near-duplicate features

Explain how SHAP values behave when A and B are near-duplicates under interventional vs conditional SHAP formulations.
Why can attributions be unstable or split unpredictably across A and B (for both GLM and GBDT)?

Defensible interpretation workflow

Propose a workflow to interpret such models: e.g., feature clustering and grouped SHAP, permutation importance conditional on the other feature, and re-fitting after removing one of the pair.
Describe the diagnostics you expect to see if A and B are redundant.

Modeling recommendations and validation

Recommend modeling changes (e.g., elastic net for GLM; feature grouping or regularization choices for trees) to handle collinearity.
Describe how you would validate that both interpretability and predictive performance remain acceptable.

Explain SHAP vs VIF under collinearity

Quick Overview

Explain SHAP vs VIF under collinearity

High Collinearity in Binary Classification: VIF, SHAP, and Interpretation Strategy

Write your answer

Explain SHAP vs VIF under collinearity

Quick Overview

Explain SHAP vs VIF under collinearity

High Collinearity in Binary Classification: VIF, SHAP, and Interpretation Strategy

Write your answer