# Stop Guessing in ML Interviews: A 5-Step Model Choice Framework

# Stop Guessing in ML Interviews: A 5-Step Model Choice Framework

<img src="https://bugfree-s3.s3.amazonaws.com/mermaid_diagrams/image_1777396573340.png" alt="Model choice framework diagram" style="width:700px;max-width:100%;height:auto;"/>

Model selection in interviews isn’t about lucky guesses — it’s about clear reasoning. Use this compact, repeatable 5-step framework to choose a model, justify it, and show interviewers you understand trade-offs.

## 1) Define the task
Start by naming the problem clearly. Common categories:

- Classification (binary / multiclass)
- Regression
- Clustering
- Anomaly detection
- Time-series forecasting

Why it matters: the task limits which algorithms and evaluation metrics make sense. In an interview, state the task up front: "This is a binary classification problem (fraud vs. not fraud)."

## 2) Read the data
Quickly summarize the dataset and call out red flags:

- Size: number of rows — tiny (<1k), moderate, large (100k+)
- Feature types: numeric, categorical, text, images
- Missing values and outliers
- Class balance for classification problems
- Label quality/noise

Interview tip: verbalize constraints: "We have ~5k samples, mostly tabular with some missingness — that suggests simpler models or careful regularization."

## 3) Match model complexity to the data
Choose models aligned with data quantity, noise, and problem complexity:

- Small / clean datasets: linear models, logistic regression, decision trees with pruning
- Moderate datasets with nonlinearity: tree ensembles (random forest, gradient boosting)
- Large datasets or complex feature types (images, text): deep learning
- High interpretability requirement: linear models, decision trees, or explainable boosting machines

Always mention compute/time constraints: "If we need a quick prototype or limited compute, I’d start with logistic regression or a light GBM."

## 4) Pick the right metric — and call out pitfalls
Choose a metric that aligns with business goals and the data:

- Accuracy: OK for balanced classes, but misleading with imbalance
- Precision / Recall / F1: use when false positives vs false negatives matter
- ROC-AUC: good general ranking metric; can be misleading on highly imbalanced datasets
- Precision@K or average precision: useful for top-k/budgeted actions
- MSE / RMSE / MAE: for regression (MAE is more robust to outliers)

In an interview, justify the metric: "We should prioritize recall because missing a positive has high cost."

## 5) Justify trade-offs and propose alternatives
Explain why you chose a model and what you’d try next:

- Accuracy vs interpretability: "I pick logistic regression for interpretability; if performance lags, I’d try XGBoost and use SHAP for explanations."
- Speed vs performance: "A light GBM is a good balance for production latency; if latency isn’t critical, I’d consider a larger neural net."
- Overfitting vs bias: "If we see high variance, I’ll add regularization or use an ensemble; if underfitting, increase model capacity or add features."

Also present fallback plans: cross-validation strategy, feature engineering ideas, and monitoring approach after deployment.

## Quick interview-ready templates
Keep these short lines ready to say when asked "Why this model?":

- "Given the dataset is small and the business needs explainability, I’d start with logistic regression." 
- "We have non-linear relationships and moderate data; a tree-based ensemble like XGBoost is a strong first choice." 
- "For high-dimensional text data with lots of labeled examples, I’d consider a transformer-based model or an LSTM depending on sequence length." 

## Common pitfalls to call out
- Choosing a complex model without enough data
- Optimizing accuracy on imbalanced classes
- Ignoring label quality
- Forgetting latency/computation constraints for production

## One-minute checklist to state in interviews
- Task: [classification/regression/...]
- Data: [size, types, missingness]
- Model choice: [model + one-sentence justification]
- Metric: [primary metric and why]
- Trade-offs: [what you accept and one alternative]

If you can succinctly explain "why this model," the interviewer knows you can reason beyond memorized names.

Happy practicing — explain your choices, not just the algorithm name.

#MachineLearning #DataScience #TechInterviews