Skip to content

Data Science Last-Minute Revision

Quickly revise the questions you picked. Add questions from the interview-questions lists with Save to → Last-Minute Revision.

30 quick revision notes for Data Science

  1. Bias is error from wrong assumptions; variance is error from sensitivity to training data; total error balances both plus noise.
  2. Data leakage means information unavailable at prediction time sneaks into training, giving falsely high validation scores.
  3. Precision is TP divided by predicted positives; recall is TP divided by actual positives; F1 is their harmonic mean.
  4. ROC AUC measures ranking quality across thresholds; prefer PR AUC when the positive class is rare.
  5. A p-value is the probability of data at least this extreme if the null hypothesis were true, not the chance the null is true.
  6. Type I error is a false positive controlled by alpha; Type II error is a false negative, and power equals one minus beta.
  7. Standard error of the mean equals the standard deviation divided by the square root of the sample size.
  8. The Central Limit Theorem says sample means are roughly normal for large samples, whatever the shape of the population.
  9. Peeking at A/B test results and stopping early inflates false positives; fix sample size and duration in advance.
  10. L1 regularisation can shrink coefficients to exactly zero for feature selection; L2 shrinks them smoothly toward zero.
  11. Bagging trains models in parallel on bootstrap samples to cut variance; boosting trains sequentially on errors to cut bias.
  12. Gradient boosting fits each new tree to the negative gradient of the loss, scaled by a learning rate.
  13. Use stratified splits for imbalanced classes and time-based splits for time series; never shuffle future data into training.
  14. Fit scalers, imputers and encoders on training data only, ideally inside a scikit-learn Pipeline.
  15. Target encoding needs out-of-fold computation and smoothing, otherwise it leaks the label into features.
  16. Tree models do not need feature scaling; KNN, SVM, PCA, linear models and neural networks usually do.
  17. Pandas merge with validate set to many_to_one catches duplicate keys that silently multiply rows.
  18. SQL logical order is FROM, WHERE, GROUP BY, HAVING, SELECT, ORDER BY, LIMIT, so SELECT aliases cannot be used in WHERE.
  19. ROW_NUMBER gives unique numbers, RANK leaves gaps after ties, and DENSE_RANK gives ties the same rank without gaps.
  20. Any comparison with NULL is unknown; use IS NULL, and remember NOT IN fails if the subquery returns a NULL.
  21. Cross-entropy loss pairs with sigmoid or softmax outputs and gives strong gradients when predictions are confidently wrong.
  22. Self-attention lets every token weigh every other token; positional encoding supplies the word order attention lacks.
  23. BERT is an encoder trained with masked language modelling; GPT-style models are decoders trained to predict the next token.
  24. RAG retrieves relevant chunks with embeddings and adds them to the prompt so the LLM answers from current, private data.
  25. Use RAG to add knowledge and fine-tuning to change behaviour or format; LoRA fine-tunes small adapter matrices only.
  26. Low temperature gives focused, repeatable LLM output; high temperature gives more varied, creative output.
  27. Training-serving skew happens when features are computed differently or with different data at training and prediction time.
  28. Data drift is a change in input distributions; concept drift is a change in the relationship between inputs and target.
  29. PSI below 0.1 means stable, 0.1 to 0.25 means monitor, and above 0.25 signals significant feature drift.
  30. Shadow deployment logs new model predictions without using them; canary sends a small share of real traffic to the new model.

Night-before checklist

  • Revise your saved questions on this page once, aloud.
  • Read the quick notes and mark any you cannot explain in one line.
  • Rehearse a two-minute self-introduction and two project stories.
  • Keep CV, ID, laptop charger and interview link ready; sleep on time.

Frequently asked questions

How does Last-Minute Revision work?

On any question page, open Save to → Last-Minute Revision. Saved questions appear here with their answers, so the night before an interview you revise only what you found hard.

Is my list saved?

Yes, in this browser. Guests can save 5 questions; free accounts can save 20 for revision and 100 for the 7-day plan.

What should I revise last before a Data Science interview?

Your saved misses first, then the quick notes on this page, then one or two project stories. Avoid learning brand-new topics the night before.