Question 1

What is the difference between supervised and unsupervised learning?

Accepted Answer

What they're testing: Whether you understand how algorithms learn and the kinds of problems they're suited for.

Answer: Supervised learning uses labeled data, meaning the model learns by example. Think of it like a teacher providing correct answers during training; the model maps input to a known output.
Unsupervised learning works with unlabeled data. The model tries to discover patterns or groupings on its own, without guidance.

Examples:
- Supervised: spam detection, fraud prediction, churn forecasting
- Unsupervised: customer segmentation, anomaly detection, topic modeling

Pro tip: Be ready to mention semi-supervised and reinforcement learning if asked about edge cases. If you want the fuller conceptual map before your interview, it's worth reviewing core machine learning algorithms so you can speak fluently across paradigms, not just define them in isolation.

Question 2

Explain overfitting and underfitting. How do you fix them?

Accepted Answer

What they're testing: Your ability to diagnose ML performance issues and improve generalization.

Answer: Overfitting: the model learns the training data too well, including noise. It performs well on the training set but poorly on unseen data.
Underfitting: the model is too simplistic to capture underlying patterns. It performs poorly on both training and test sets.

How to fix overfitting: regularization (L1/L2), simplifying the model, pruning trees, adding dropout (for deep learning), or getting more data.
How to fix underfitting: a more complex model, reducing regularization, increasing training time, or improving feature engineering.

Real-world analogy: overfitting is like memorizing answers without understanding concepts. Underfitting is like guessing because you didn't study.

Question 3

How does a decision tree work? How can it be improved?

Accepted Answer

What they're testing: Understanding of classical ML models and their limitations.

Answer: A decision tree splits data into branches based on feature thresholds that reduce impurity. At each node, it picks the best feature to split on using criteria like Gini impurity or entropy/information gain. It continues splitting until it hits a stopping condition, like max depth or minimum samples.

Weaknesses: prone to overfitting, and sensitive to small changes in data.

How to improve:
- Pruning: removing branches that add little value
- Ensemble methods: Random Forest (bagging) averages multiple trees to reduce variance; Gradient Boosting (boosting) builds trees sequentially to correct previous errors

Pro tip: If the interviewer digs deeper, talk about feature importance and how decision trees support interpretability, a theme that comes up again later in this list.

Question 4

What's the difference between bagging and boosting?

Accepted Answer

What they're testing: Your grasp of ensemble methods and when to use which.

Answer:
Bagging (Bootstrap Aggregating) trains multiple models in parallel on random subsets of the data (with replacement), then combines their outputs, usually by averaging (regression) or majority voting (classification). Goal: reduce variance and prevent overfitting.
Boosting trains models sequentially. Each new model tries to fix the errors of the previous one. Goal: reduce bias and build a strong learner from many weak learners.

Pro tip: In interviews, say you prefer Random Forests for quick baseline models and Boosting (like XGBoost or LightGBM) for performance tuned pipelines in competitions or production.

Question 5

Walk me through the lifecycle of a machine learning project you've handled.

Accepted Answer

What they're testing: Whether you understand the real-world process beyond model training.

Answer structure (STAR format):
- Situation: the business problem
- Task: your objective
- Action: step-by-step actions across the ML lifecycle
- Result: outcomes and impact (metrics, revenue, adoption)

Example answer: "At my internship, the goal was to reduce customer churn using predictive analytics. We started with stakeholder interviews to understand KPIs, then moved into data collection from CRM systems. After cleaning and feature engineering, I trained multiple models and settled on a gradient boosting classifier with an 82% F1-score. We integrated the model into a dashboard that helped the customer success team proactively engage at-risk users, reducing churn by 12% over three months."

Bonus tip: Mention tools, Python (Pandas, Scikit-learn), MLflow for experiment tracking, or Docker for deployment. Recent field data on 2026 interview loops shows a strong, consistent pattern here: interviewers increasingly ask candidates to walk through a system they actually improved and what broke in production, not a hypothetical. Having a real, specific project story, backed by numbers, is one of

Question 6

How do you handle imbalanced datasets?

Accepted Answer

What they're testing: How you think about fairness, recall, and minority-class handling.

Answer: When one class dominates the dataset, say 95% negative, 5% positive, your model might predict only the majority class and still appear "accurate."

Techniques to handle imbalance:
- Resampling: oversampling (SMOTE) or undersampling the majority class
- Algorithmic approaches: class weight adjustment, or anomaly detection algorithms if the minority class is rare but critical
- Evaluation metric shifts: precision, recall, F1-score, or ROC-AUC instead of accuracy
- Domain-specific strategies: custom loss functions like focal loss for high class imbalance in deep learning

Pro tip: In real-world use cases like fraud detection or medical diagnosis, be ready to explain why you'd rather tolerate a few more false positives than miss a true anomaly. That trade-off is exactly what the next question digs into.

Question 7

What evaluation metrics do you use and when?

Accepted Answer

What they're testing: Whether you know when accuracy isn't enough, and how to justify your metric choice based on the problem type.

Answer: Choosing the right evaluation metric depends on the business objective and the data's characteristics.
- Accuracy: useful when classes are balanced
- Precision: of the predicted positives, how many are correct (good for reducing false positives)
- Recall: of the actual positives, how many did you catch (good for minimizing false negatives)
- F1 score: harmonic mean of precision and recall, strong choice for imbalanced data
- ROC-AUC: the probability the classifier ranks a random positive higher than a random negative, useful for probabilistic models

Real-world example: In a fraud detection project, accuracy was a poor fit because fraud was only 2% of transactions. The focus shifted to recall, since missing a fraudulent case was costly, and the model was optimized using F1 and ROC-AUC instead.

Pro tip: Always relate your metric choice back to business risk. For a deeper, worked breakdown of exactly this metric with real examples, the F1 score in machine learning is worth reviewing before your interview.

Question 8

What are the key components of a production ML pipeline?

Accepted Answer

What they're testing: Whether you think like a developer who can ship production-grade systems, not just a data scientist who stops at the notebook.

Answer: A robust ML pipeline includes:
- Data ingestion, pulling from databases, APIs, logs, or streaming sources
- Data validation and cleaning, detecting anomalies, missing values, and schema changes
- Feature engineering, real-time or batch, with consistency across train, test, and live environments
- Model training and versioning, automating experiments and tracking metrics with tools like MLflow or DVC
- Model serving, REST APIs (Flask, FastAPI), batch jobs, or edge deployment
- Monitoring and retraining, tracking data drift, concept drift, model decay, and latency

Tools worth naming: Airflow, TFX, SageMaker, Kubeflow, Docker, Prometheus, Grafana.

Bonus tip: If you've worked on CI/CD for ML models, bring it up, it's a genuine differentiator. This question has only gotten more central to 2026 hiring loops. Multiple recent analyses of ML engineering job postings note that the market isn't short on candidates who know algorithms; it's short on engineers who've actually run something in production and can speak to what broke.

Question 9

Explain gradient descent and its variants.

Accepted Answer

What they're testing: Whether you deeply understand optimization, the mechanism at the heart of model training.

Answer: Gradient descent is an optimization algorithm that minimizes a cost function by updating parameters in the direction of the negative gradient.

Variants:
- Batch gradient descent: uses the entire dataset, stable but slow and memory-intensive
- Stochastic gradient descent (SGD): updates weights after each training example, faster but noisier
- Mini-batch gradient descent: updates on small batches (e.g., 32 samples), the standard approach in deep learning
- Adam: combines momentum and adaptive learning rates, fast convergence, widely used by default
- RMSprop, Adagrad: adaptive learning rate optimizers suited to sparse data or noisy gradients

Analogy: finding the lowest point in a foggy valley using small, corrective steps.

Pro tip: If pushed further, be ready to explain learning rate decay, exploding or vanishing gradients, or how you chose an optimizer in a past project. Understanding how these optimization mechanics differ from the deep learning models built on top of them is part of a broader distinction worth having clear in your head: deep learning vs machi

Question 10

How do you ensure model interpretability?

Accepted Answer

What they're testing: Your awareness of explainable AI (XAI), especially relevant in regulated industries.

Answer: Interpretability means understanding why a model made a specific prediction.

Approaches:
- Model choice: favor interpretable models (decision trees, linear or logistic regression) when transparency is critical
- Post-hoc methods: SHAP (SHapley Additive exPlanations) for global and local feature contributions; LIME (Local Interpretable Model-agnostic Explanations), which perturbs input to study output changes; partial dependence plots and feature importance

Use cases: finance (credit scoring), healthcare (disease risk prediction), legal (recidivism prediction).

Pro tip: Talk about the trade-off between accuracy and explainability, especially relevant for models like deep neural networks, where the most accurate option is often the least interpretable one.

Must Do

Must Do