Apple·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say

Apr 2026

Summary

Apple data scientist interview with a meaty modeling critique question. No fluff, just a wall of statistical concepts thrown at you and expected to untangle them coherently. Felt more like a written exam than a conversation.

Questions Asked (1)

You're reviewing a linear regression workflow built by another data scientist for modeling session dwell time on a website. The analyst assumed OLS was valid because Y looked normally distributed, fit every possible predictor combination including squared and interaction terms, and picked the model with the best in-sample fit. Critique the workflow: what would you clarify, what's wrong with each step, and how would you redesign it?

Data ModelingTechnical Trade-offsProduct Analytics & Metrics

Author's notes

This question is basically five questions stapled together, which I didn't fully appreciate until I was already three minutes into talking about Gauss-Markov and realized I hadn't touched overfitting yet.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business goal and data context, then systematically critique each step of the workflow, and finally propose a redesigned approach that emphasizes out-of-sample validation, appropriate model assumptions, and interpretability.

Pro tip: Frame your critique around the bias-variance trade-off and the importance of aligning the model with the decision it will inform—this shows you think like a product data scientist, not just a statistician.

1. Clarify the objective and data

Ask what decision the model will support, what the target variable represents, and how the data was collected. This ensures the modeling approach aligns with business needs.

2. Critique the OLS assumption check

Explain that normality of Y is not required for OLS; instead, focus on linearity, independence, homoscedasticity, and normality of residuals. Also consider if dwell time is skewed or zero-inflated.

3. Critique the model selection process

Point out that fitting all possible predictor combinations with squared and interaction terms leads to overfitting and multicollinearity. In-sample fit is not a valid measure of model performance.

4. Propose a redesigned workflow

Suggest using regularization (e.g., LASSO), cross-validation for model selection, and checking out-of-sample performance. Consider transformations or alternative models (e.g., gamma regression) if assumptions are violated.

5. Emphasize validation and interpretability

Highlight the need for holdout or cross-validation, residual diagnostics, and ensuring the final model is interpretable and actionable for stakeholders.

Key Points to Mention

OLS assumptions: normality of residuals, not Y; homoscedasticity; linearity; independence
Overfitting due to exhaustive search and in-sample evaluation
Multicollinearity from squared and interaction terms without centering
Use of cross-validation and regularization (LASSO, ridge) for model selection
Alternative models for skewed positive continuous data (e.g., gamma, log-normal)
Business context: dwell time may have outliers, zero-inflation, or nonlinear relationships

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.