I rattled off multicollinearity and heteroscedasticity and then kind of stalled.
Start by acknowledging that linear regression is a foundational model but has several assumptions and limitations that can lead to failure modes. Then, systematically categorize the problems into data-related, model-related, and evaluation-related issues, and briefly mention how to detect and mitigate each. Conclude by tying it back to practical software engineering contexts, such as when building ML pipelines or debugging models.
Pro tip: Demonstrate maturity by not just listing problems but also explaining how you would diagnose them (e.g., residual plots for heteroscedasticity) and the trade-offs of fixes (e.g., regularization vs. feature selection). This shows you can apply the knowledge in real-world engineering decisions.
Discuss how linear regression assumes linearity, independence, homoscedasticity, and normality of residuals. Violations lead to biased or inefficient estimates.
Explain that highly correlated predictors inflate coefficient variance, making interpretation unreliable and causing overfitting.
Cover how too many features (relative to samples) cause overfitting, while too few or overly simplistic models underfit. Mention regularization techniques like Ridge/Lasso.
Highlight that outliers can disproportionately affect the fit, leading to poor generalization. Discuss detection via leverage and Cook's distance.
Address missing values, measurement errors, and irrelevant features that degrade model performance. Emphasize preprocessing steps.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.