Machine Learning Fundamentals Questions

Core machine learning concepts and terminology for conceptual understanding. Topics include supervised and unsupervised learning, regression and classification problems, training validation and test splits, cross validation, loss functions and optimization at a high level, model evaluation metrics, overfitting and underfitting, regularization concepts, and common basic model families such as linear models decision trees nearest neighbors and simple neural networks. Emphasis is on conceptual explanations and trade offs rather than deep mathematical derivations.

HardTechnical

95 practiced

Your company wants to ensure a new model meets basic fairness and responsible-AI standards. At a practical, non-mathematical level outline the steps an ML Engineer should take: define fairness goals with stakeholders, select appropriate fairness metrics, collect subgroup performance data, apply mitigation strategies (pre-, in-, post-processing), and set up monitoring and audit logs for ongoing assessment.

Sample Answer

Situation: We're building a model that impacts people (e.g., loan approvals). As the ML Engineer, I lead the technical implementation of fairness and responsible-AI practices end-to-end.

1) Define fairness goals with stakeholders- Convene product, legal, ethics, ops and a representative user advocate group.- Translate business objectives into clear constraints: which protected attributes matter, acceptable trade-offs (e.g., small accuracy loss for large fairness gain), regulatory requirements, and user experience limits.- Deliverable: a short charter with success criteria (business and fairness KPIs) and decision owners.

2) Select appropriate fairness metrics- Map goals to practical metrics (e.g., demographic parity, equal opportunity, disparate impact) and operational metrics (false positive/negative rates by group, calibration).- Choose a small set (primary + secondary) and explain trade-offs to stakeholders in plain language.- Deliverable: metric spec with measurement frequency and thresholds.

3) Collect subgroup performance data- Ensure data pipelines capture self-identified protected attributes or proxies ethically and legally; where unavailable, use careful sampling and qualitative review.- Instrument model evaluation to slice performance by subgroup, intersectional groups, and operational cohorts (geography, device, etc.).- Deliverable: reproducible evaluation notebooks and dashboards that show subgroup metrics.

4) Apply mitigation strategies (pre-, in-, post-processing)- Pre-processing: rebalance training data, label auditing, feature engineering to remove leaky proxies.- In-processing: add fairness-aware objectives or constraints, use calibrated thresholds per group if justified.- Post-processing: adjust scores/thresholds, human-in-the-loop overrides for borderline cases.- Run A/B tests and human reviews to verify user impact before full rollout.- Deliverable: comparison of candidate mitigations, chosen approach, and rollback plan.

5) Monitoring, audit logs and governance- Deploy monitoring: real-time dashboards for subgroup metrics, drift detection for inputs and outputs, alerting when thresholds exceeded.- Maintain immutable audit logs: inputs (or hashes), model version, decision, confidence, and downstream actions for traceability.- Schedule periodic audits (monthly for key metrics, quarterly ethical review) and ad-hoc investigations on incidents.- Build playbooks: how to triage alerts, notify stakeholders, and enact mitigations/rollbacks.

Example outcome: For a lending model I led, we defined equal opportunity as the primary fairness goal, measured false negative rates by race and income, rebalanced training data and applied group-specific thresholds, then monitored production with alerts. That approach reduced subgroup false negatives by 18% with a 1.5% overall accuracy drop, and provided clear audit trails for compliance.

Key leadership points:- Make decisions transparent and documented.- Balance technical fixes with product impact and legal constraints.- Ship incrementally: validate mitigations experimentally, then scale with monitoring and governance.

MediumTechnical

85 practiced

Explain how to choose a loss function for regression problems. Compare MSE, MAE, Huber loss, and quantile loss in terms of robustness to outliers, optimization behavior, and business use cases where each is preferable.

Sample Answer

When choosing a regression loss, match the loss’s statistical assumptions, optimization properties, and business objective (e.g., penalize large errors, robustness, or asymmetric costs).

Quick definitions:- MSE (mean squared error): (1/n) Σ (y - ŷ)² — emphasizes large errors (quadratic).- MAE (mean absolute error): (1/n) Σ |y - ŷ| — linear penalty, robust to outliers.- Huber loss: quadratic near zero and linear beyond threshold δ — trades off sensitivity and robustness.- Quantile loss (pinball): for quantile τ, penalizes under/over-predictions asymmetrically — useful for conditional quantiles.

Comparison:- Robustness to outliers: MAE and quantile losses are more robust than MSE. Huber is tunable (with δ) to be robust while retaining smoothness. MSE is least robust.- Optimization behavior: MSE is smooth and convex with well-behaved gradients (fast convergence with gradient methods). MAE has nondifferentiable point at zero (subgradient methods or smoothing needed), which can slow optimization. Huber is smooth everywhere and often gives stable gradients. Quantile loss is convex but nondifferentiable at the quantile breakpoint; optimization typically via linear programming or subgradient methods.- Statistical interpretation: MSE corresponds to maximizing likelihood under Gaussian noise; MAE to Laplace noise; quantile to asymmetric Laplace/conditional quantiles.- Business use cases: - MSE: when large errors must be heavily penalized (e.g., forecasting for capacity planning where big misses are costly) and data has few outliers. - MAE: when you want median behavior and robustness (e.g., typical demand forecasting with noisy outliers). - Huber: when most errors are small but occasional outliers exist — good default for regression in production. - Quantile loss: when you need prediction intervals or asymmetric cost sensitivity (e.g., risk management, inventory safety stock where underprediction is costlier).

Practical tips:- Inspect residual distribution and business cost asymmetry first.- Use Huber as a robust, optimizer-friendly default; tune δ by cross-validation.- For intervals, train multiple quantiles (e.g., 0.1, 0.5, 0.9) to build prediction bands.

HardSystem Design

68 practiced

You're designing a retraining cadence for a classification model. Describe factors that should influence retraining frequency (data volume, drift detection, cost, SLA), propose a hybrid schedule combining periodic and event-driven retraining, and explain how you'd validate retrained models before swap-in.

HardTechnical

94 practiced

Explain bagging, boosting, and stacking ensemble techniques at a conceptual level. For each method describe how it reduces error (bias or variance), identify common failure modes, and discuss production trade-offs such as increased latency, storage, interpretability, and complexity of model updates.

Sample Answer

Bagging (Bootstrap Aggregating)- Concept: Train many independent base learners (usually high-variance models like decision trees) on different bootstrap samples and average (or vote) their predictions.- Error reduction: Primarily reduces variance by averaging uncorrelated errors; bias stays roughly the same as each base learner.- Failure modes: If base learners are high-bias (e.g., shallow trees), bagging won’t help; if bootstrap samples aren’t diverse (small dataset, strong feature leakage), ensemble gains are limited.- Production trade-offs: Increased storage (many models), higher inference latency if models evaluated sequentially (can parallelize), lower interpretability, more complex deployment and model update pipelines (need consistent versioning and coordinated retraining).

Boosting- Concept: Sequentially train weak learners where each new learner focuses on examples previous learners mispredicted (e.g., AdaBoost, Gradient Boosting). Final model is weighted sum of learners.- Error reduction: Reduces bias by creating a strong learner from weak ones; can also reduce variance but can overfit if too expressive.- Failure modes: Sensitive to noisy/ mislabeled data (weights amplify noise), overfitting if too many rounds, and poor calibration without regularization.- Production trade-offs: Model size and latency grow with number of trees/estimators; often faster at inference than bagging if trees are shallow but still heavier than single model; harder to interpret; updates require careful incremental training or full retrain; feature importance available but less transparent.

Stacking (Stacked Generalization)- Concept: Train diverse base models (different algorithms/hyperparams), then train a meta-learner on base models’ out-of-fold predictions to combine strengths.- Error reduction: Can reduce both bias and variance by leveraging complementary error patterns across learners.- Failure modes: Risk of meta-learner overfitting if out-of-fold predictions aren’t properly generated; if base models are correlated, little gain; complexity increases chance of pipeline bugs.- Production trade-offs: Highest complexity — multiple model families + meta model; significant storage and latency unless distilled; updates require coordinated retraining of base and meta models; interpretability low but can inspect base contributions.

Practical guidance: choose bagging for variance control and robustness, boosting for maximizing predictive power with careful regularization, stacking when different model families provide complementary gains and you can manage added operational complexity. Consider model distillation, caching, parallel inference, and CI/CD orchestration to mitigate production costs.

HardSystem Design

73 practiced

Describe a simple end-to-end pipeline to take tabular data from raw logs to a deployed binary classifier. Include steps for data ingestion, feature engineering, training, validation, deployment, monitoring, and retraining triggers. Keep the description high-level and practical for a small engineering team.

Unlock Full Question Bank

Get access to hundreds of Machine Learning Fundamentals interview questions and detailed answers.

Join thousands of developers preparing for their dream job.