Machine Learning Engineer interview questions

For squared error, expected prediction error can be viewed as bias squared, variance, and irreducible noise.

• Bias measures systematic error from assumptions that are too restrictive, such as fitting a linear model to a strongly nonlinear relationship.
• Variance measures how much the fitted model would change if it were trained on another sample from the same population.
• Increasing model flexibility often lowers bias but raises variance, so validation estimates the balance that generalizes best.
• Irreducible noise comes from randomness or missing information and cannot be removed by choosing a more complex model.

Empirical risk is average loss on observed samples, while expected risk is average loss over the unknown data distribution.

• Training directly minimizes empirical risk because the true population distribution is unavailable.
• A flexible model can drive empirical risk down without reducing expected risk if it fits sample-specific noise.
• Validation data estimates expected risk on unseen observations drawn under similar conditions.
• Regularization and capacity control restrict the solutions considered during empirical risk minimization.

L1 and L2 add different weight penalties to the data loss, changing which parameter solutions the optimizer prefers.

• L1 adds the sum of absolute coefficients and has a sharp corner at zero, so some coefficients can become exactly zero.
• L2 adds the sum of squared coefficients and continuously shrinks large weights without usually eliminating them.
• The regularization strength trades training fit for lower complexity and is selected on validation data.
• Features should usually be scaled first because the penalty acts directly on coefficient magnitude.

Elastic net combines L1 sparsity with L2 stability and is useful when many predictors are correlated.

• Pure L1 may select one feature from a correlated group and discard the others unpredictably.
• The L2 component encourages correlated predictors to enter or leave the model more smoothly as a group.
• The L1 component can still remove weak features and produce a compact model.
• Both total penalty strength and the L1-to-L2 ratio require validation-based tuning.

Early stopping limits effective model capacity by ending optimization before the model fits training noise.

• Training loss may keep falling after validation loss starts rising, which signals deteriorating generalization.
• The best checkpoint is selected by a validation metric with a patience window to avoid reacting to random fluctuations.
• It is especially common for boosting and neural networks whose capacity grows with additional iterations.
• The validation set used for stopping is part of model selection and must not also serve as the final test set.

Dropout randomly masks activations during training so the network cannot depend too heavily on specific paths.

• Each training step samples a different thinned network, which discourages fragile co-adaptations between units.
• At inference, all units are used with scaling handled so expected activation magnitudes remain consistent.
• A higher dropout rate increases regularization but can cause underfitting when too much signal is removed.
• Dropout is less universally helpful in architectures that already use strong normalization, data augmentation, or large datasets.

Batch normalization standardizes intermediate activations using batch statistics and then learns a scale and shift.

• It can smooth optimization and permit larger learning rates by keeping activation distributions better conditioned.
• Training uses statistics from each mini-batch, while inference uses accumulated running statistics.
• Very small or nonrepresentative batches can make those estimates noisy and reduce effectiveness.
• Its stochastic batch statistics can add some regularization, but normalization is primarily an optimization aid.

A calibrated classifier assigns probabilities that match observed outcome frequencies.

• Among predictions near 0.8, roughly 80 percent should be positive when calibration is good.
• Ranking metrics such as ROC-AUC can be strong even when predicted probabilities are poorly calibrated.
• Reliability diagrams and the Brier score help assess calibration across probability ranges.
• Platt scaling or isotonic regression can recalibrate scores using held-out data.

A classification threshold should reflect error costs, class prevalence, and the operational objective rather than defaulting to 0.5.

• Lowering the threshold usually raises recall while increasing false positives.
• Raising it usually improves precision while missing more true positives.
• The threshold must be chosen on validation data with a metric or cost function tied to the intended decision.
• Probability calibration matters when the threshold represents an explicit risk or expected-value boundary.

Both are proper scoring rules that reward accurate probabilities, but they penalize errors differently.

• Log loss uses the negative log probability of the true class and heavily punishes confident incorrect predictions.
• The Brier score is mean squared error between predicted probabilities and binary outcomes.
• Both assess information that accuracy discards when it reduces probabilities to hard labels.
• Scores should be compared against a prevalence-based baseline and interpreted with class balance in mind.

A precision-recall curve is often more informative when the positive class is rare and positive performance is the main concern.

• ROC uses false positive rate, whose denominator includes many negatives and can make performance look strong under severe imbalance.
• Precision directly reflects how many predicted positives are correct, so it exposes false-positive burden.
• Recall shows how much of the positive class is recovered as the threshold changes.
• PR-AUC depends on prevalence, so comparisons require datasets with comparable class distributions.

Macro, micro, and weighted averages combine per-class results with different priorities.

• Macro averaging gives every class equal weight, so rare-class performance has the same influence as common-class performance.
• Micro averaging pools all decisions before computing the metric, which emphasizes frequent classes.
• Weighted averaging takes the per-class metric weighted by class support and can hide weak minority classes.
• The appropriate average depends on whether each class, each example, or observed prevalence should drive the result.

The number of folds and dependence between splits determine much of a cross-validation estimate's bias, variance, and cost.

• More folds train on a larger fraction of data, reducing pessimistic bias but increasing compute.
• Fold scores are correlated because their training sets overlap, so their standard deviation is not a simple confidence interval.
• Repeated cross-validation reveals sensitivity to random partitioning but does not fix a mismatched splitting strategy.
• Grouped, temporal, or stratified structure matters more than choosing a conventional value of k.

Nested cross-validation separates hyperparameter selection from performance estimation when data is limited.

• The inner loop tunes hyperparameters using only the outer training portion.
• The outer loop evaluates the entire selection procedure on data untouched by the inner loop.
• Reporting the best inner score is optimistically biased because many configurations were compared.
• Nested validation costs substantially more compute but gives a cleaner estimate before final retraining.

Grouped cross-validation is required when observations from the same entity are dependent and must stay in one fold.

• Rows from one user, patient, device, or document source can share patterns that a model could memorize.
• Random row splitting leaks those entity-specific patterns into both training and validation data.
• GroupKFold keeps each group entirely on one side of a split while rotating held-out groups.
• Stratified group methods may be needed when both group isolation and class balance matter.

Walk-forward validation preserves time order by training on the past and evaluating on a later interval.

• Expanding windows keep all earlier observations, while rolling windows keep a fixed amount of recent history.
• Multiple cutoffs reveal whether quality changes across different market, seasonal, or behavioral periods.
• Any feature, target, and preprocessing statistic must be available as of each historical prediction time.
• Random folds are inappropriate when future observations can reveal information about earlier ones.

Preprocessing leaks information when a transformation learns from validation rows before those rows are evaluated.

• Scaling on the full dataset exposes validation means and variances to the training process.
• Imputation, feature selection, PCA, and target encoding can leak for the same reason.
• Each transformation must be fitted inside the training portion of every fold and only applied to its validation portion.
• A scikit-learn Pipeline enforces this order when passed to cross-validation or search utilities.

MCAR, MAR, and MNAR describe what the probability of a missing value depends on.

• MCAR means missingness is unrelated to observed or unobserved values, which is a strong and uncommon assumption.
• MAR means missingness can be explained by other observed variables, making conditional imputation more defensible.
• MNAR means missingness depends on the missing value itself or another unobserved cause.
• No imputation method can fully resolve MNAR from the observed table alone without assumptions or additional data.

Target encoding replaces a category with a target statistic, producing compact signal but creating serious leakage risk.

• Rare categories have noisy estimates, so smoothing toward the global mean reduces variance.
• Training rows need out-of-fold or leave-one-out encodings so their own targets do not define their features.
• Unseen categories require a fallback such as the global mean or a learned prior.
• The encoding must be recomputed within every validation split and respect time order when data is temporal.

Interaction and polynomial features let a simple model represent relationships beyond independent linear effects.

• An interaction term allows one feature's effect to depend on another feature's value.
• Polynomial terms can approximate curved relationships while retaining a linear estimator in the expanded feature space.
• Expansion grows dimensionality rapidly and can increase multicollinearity, variance, and computation.
• Scaling, regularization, and domain-guided selection keep the expanded representation manageable.