These 64 Data Science questions answer common beginner questions about Statistics, Machine Learning and Model Monitoring. Use the topic links to find a short explanation, then try the practice exercise to apply it.
Explore the ScholarLogic Data Science course in Bengaluru, based in BTM Layout, or try the beginner practice exercise.
Statistics and data
What is a population in statistics?
It is the complete set you want to describe. Define its people, events or time period before choosing a sample.
What is sampling bias?
It occurs when the way data is collected systematically excludes or overrepresents parts of the intended population.
When is the median preferable to the mean?
It is useful for skewed values such as incomes or response times. Report the mean too if total impact matters.
What does standard deviation describe?
It measures spread around the mean in the same units as the values. A few extreme observations can increase it.
What is the interquartile range?
It is the difference between the 75th and 25th percentiles, describing the spread of the middle half.
What is an empirical distribution?
It describes the values observed in a sample and their frequencies. It may differ from the underlying population distribution.
What is conditional probability?
It is the probability of an event given another event has occurred. The condition determines the relevant denominator.
What does independence mean in probability?
Knowing one event occurred does not change the probability of the other, where the conditional probability is defined.
What is a confidence interval?
It is produced by a method designed to cover a fixed parameter at a stated rate across repeated samples under its assumptions.
What is a p-value?
Under the null hypothesis and model assumptions, it is the probability of a result at least as extreme as the observed one.
What is a false positive in hypothesis testing?
It means rejecting a true null hypothesis. The significance level controls this risk under the test assumptions.
Why report effect size alongside statistical significance?
Effect size describes how large the observed difference is. A small effect can be statistically detectable without being useful.
What is a confounder?
It is a factor related to both an exposure and an outcome that can distort their observed association.
How can Simpson’s paradox change a conclusion?
An overall pattern can reverse within groups when the groups have different compositions. Inspect group rates and sizes.
What does correlation measure?
It measures association. Pearson correlation measures linear association and can be influenced by outliers.
What is bootstrap resampling?
It repeatedly samples observations with replacement from the observed data. Adapt the method for dependent observations.
Data preparation
What is a data dictionary in a machine learning project?
It records field meanings, types, units, allowable values and availability so someone can reconstruct inputs.
How do I detect an impossible timestamp in training data?
Check allowed ranges and event ordering, then inspect timezone conversion and source errors.
What does missing not at random mean?
Missingness can depend on the unobserved value even after accounting for observed data. A simple imputation can bias results.
When is median imputation reasonable?
It can be a simple numeric baseline when appropriate. Fit it on training data and test how missingness affects results.
What is one-hot encoding?
It represents categories with indicator columns. Plan for previously unseen categories and high cardinality.
Why is target encoding easy to misuse?
It uses outcomes to encode categories. Estimate encodings without letting a row’s own outcome or held-out outcomes leak into its features.
When should I group rare categories?
Consider it when estimates for individual categories are unstable. Learn the grouping from training data and retain meaningful distinctions.
What is a prediction-time feature check?
For each input, verify that it was actually available at the moment the model is meant to make its prediction.
Why can repeated customers leak across a split?
Rows for one customer may share behavior or identifiers. Group by customer when evaluating generalization to new customers.
What is a chronological backtest?
It repeatedly trains on past observations and tests on a later window, simulating what would have been known at the time.
Why record dataset versions?
Version records let reviewers connect a result to the precise data and transformations used.
What is a reproducible feature transformation?
It has documented inputs and consistent behavior between training and deployment, including for missing or unfamiliar values.
Why inspect distributions by source system?
A change in source can explain shifts in values or missing fields that otherwise look like a change in customer behavior.
How do I handle an unknown category at prediction time?
Define a fallback or rejection policy and test it. Avoid silently treating unknown values as a known category.
What is data lineage for a feature?
It traces the feature from its original fields through transformations, ownership and timing.
How do I detect duplicate observations before modelling?
Define what counts as the same observation and look for repeated keys or records. Resolve genuine repeats according to the task.
Models and interpretation
What is logistic regression used for?
It models a class probability, commonly for a binary outcome. Apply a decision threshold chosen for the task.
What does a coefficient mean in linear regression?
It is the estimated change in prediction per unit of one feature, holding the other model features fixed.
What is multicollinearity?
It occurs when predictors are strongly related. It can make individual coefficient estimates unstable.
How does k-nearest neighbors classify a point?
It uses the labels of nearby training examples. The distance metric and scaling affect which points are neighbors.
What is k-means clustering?
It assigns points to a chosen number of clusters by minimizing within-cluster squared distances around centroids.
How does PCA reduce dimensions?
It forms linear components ordered by explained variance. Feature scaling and interpretability matter.
What is a decision boundary?
It separates inputs that a classifier assigns to different labels under a decision rule.
How does boosting combine models?
It builds learners sequentially so later learners improve the current ensemble under a chosen objective.
What is early stopping?
It stops training when a monitored validation measure no longer improves enough. Keep the final test data separate.
What is a learning curve in model development?
It plots performance against training size or training progress. Label the horizontal axis so readers know which it shows.
What is permutation feature importance?
It measures how a score changes when feature values are shuffled. Correlated inputs can complicate interpretation.
What is a partial dependence plot?
It shows how average predictions change across a feature’s values. Correlated features can create unrealistic combinations.
Why can a model explanation differ from a causal explanation?
A model describes how inputs relate to its predictions. A causal claim needs evidence about what changes under intervention.
What is a residual plot useful for?
It can reveal systematic prediction errors, changing variance or unusual observations.
How can a simple heuristic compete with a model?
Use a rule available at prediction time as a benchmark. Compare it on the same held-out rows and metrics.
What is the bias-variance trade-off?
A rigid model may miss real patterns, while an overly flexible one can be sensitive to the training sample.
Evaluation and monitoring
What does ROC AUC summarize?
It summarizes how well a score ranks positives above negatives across thresholds. It does not specify a deployment threshold.
What is balanced accuracy?
For binary classification, it is the mean of sensitivity and specificity, giving both classes equal weight.
When is a cost-sensitive evaluation useful?
Use it when error types have different consequences. Apply agreed costs and compare candidate decision rules.
How do I compare model performance across subgroups?
Report the same metrics and sample sizes per group. Show uncertainty when groups are small.
What is a calibration curve?
It compares predicted probabilities with observed event rates across groups of predictions.
Why assess performance on recent data?
The deployment population may change. A recent held-out window can reveal deterioration hidden by older results.
What is concept drift?
It is a change in the relationship between inputs and outcomes. Input drift alone does not prove it.
What is a delayed label?
It is an outcome known only after prediction, such as whether an order was returned after several days.
How can I monitor predictions before labels arrive?
Track input validity, missingness, score distributions and prediction volume, then confirm performance once labels mature.
What is a shadow deployment for a model?
The new model scores live inputs without controlling decisions. Compare its outputs and operational behavior before adoption.
When should a model fall back to a rule?
Define a fallback for unsupported inputs, unavailable dependencies or other expected failures, and monitor how often it runs.
What is a model card?
It documents intended use, training and evaluation, known limits and operational considerations.
Why track preprocessing with the model artifact?
The model expects inputs transformed consistently. Version the transformation and model so they can be deployed together.
How should I choose a model retraining trigger?
Use evidence such as label-based degradation or meaningful drift, and validate the replacement before release.
How do I make a notebook project reproducible?
List dependencies, provide data instructions and check that a clean run from top to bottom recreates the results.
What should a data science portfolio include beyond a model score?
Show the problem, baseline, split, error analysis, limits and how the prediction would be used.
For further reading, see the official reference and compare ScholarLogic’s six courses.
