Machine Learning

How to answer in an interview

Overfitting happens when a model learns the training data too well, including noise and accidental patterns, so it performs well on training data but poorly on new data. I reduce it by collecting more data, simplifying the model, using regularization, applying early stopping, tuning hyperparameters on validation data, and checking performance on a truly untouched test set. The real goal is generalization, not memorization.

Explain like I am 5

Overfitting is like memorizing the answers to one practice test instead of learning the subject. You do great on the practice test, but you struggle when the questions change.

How to answer in an interview

We split data so each part has a different role. The training set teaches the model. The validation set helps choose models, features, and hyperparameters. The test set is reserved for a final unbiased estimate of performance after all decisions are made. This separation helps prevent data leakage and gives us a more realistic view of how the model will behave on new data.

Explain like I am 5

You learn with the training set, practice your choices with the validation set, and save the test set for the final surprise check.

How to answer in an interview

Cross-validation is a technique for estimating how well a model generalizes by splitting the data into multiple folds, training on some folds, and evaluating on the remaining fold. K-fold cross-validation repeats this process k times so every data point serves as both training and validation data. It gives a more robust estimate of performance than a single train-validation split, especially when data is limited. I use it for model selection, hyperparameter tuning, and detecting overfitting before committing to a final evaluation on the held-out test set.

Explain like I am 5

Cross-validation is like practicing a test multiple times with different study groups. It tells you how well you really know the material, not just how well you did on one practice test.

How to answer in an interview

Decision trees work by recursively splitting the data on feature values that best separate the target classes or reduce prediction error. Common splitting criteria include Gini impurity, entropy, and mean squared error. They are easy to interpret and handle mixed data types, but they are prone to overfitting, sensitive to small changes in data, and can create overly complex trees. In practice, single decision trees are rarely used alone; instead, ensemble methods like random forests and gradient boosting combine many trees to reduce variance and improve accuracy.

Explain like I am 5

A decision tree is like a flowchart of yes-or-no questions that leads to a prediction. Single trees can be too detailed, so we combine many trees for better results.

How to answer in an interview

A neural network is a computational model inspired by the brain, made up of layers of interconnected nodes or neurons. Each neuron applies a weighted sum of its inputs, adds a bias, and passes the result through an activation function like ReLU, sigmoid, or softmax. The network learns by adjusting weights through backpropagation and gradient descent to minimize a loss function. Common architectures include fully connected networks, convolutional networks for images, and recurrent or transformer networks for sequences. The depth and width of the network determine its capacity to learn complex patterns.

Explain like I am 5

A neural network is like layers of filters that transform input into a prediction. Each layer learns to detect increasingly complex patterns.

How to answer in an interview

Parametric models assume a fixed functional form with a finite number of parameters that are learned from data, such as linear regression or logistic regression. They are simple, interpretable, and fast to train, but they may underfit if the true relationship is complex. Non-parametric models make fewer assumptions about the underlying form and can adapt their complexity to the data, such as k-nearest neighbors, decision trees, and kernel methods. They are more flexible but often require more data and computation. The choice depends on the amount of data, the complexity of the relationship, and the need for interpretability.

Explain like I am 5

Parametric models make a simple assumption about what the answer looks like. Non-parametric models are more flexible and let the data shape the answer without a fixed template.

How to answer in an interview

I care more about precision when false positives are expensive, because I want the model to be very reliable when it predicts a positive. I care more about recall when missing true positives is worse than raising some extra false alarms. For example, fraud blocking and medical screening often have different trade-offs. The best metric depends on the business cost of false positives versus false negatives, and I usually pair the metric choice with a threshold analysis.

Explain like I am 5

Precision is about being careful when you say yes. Recall is about finding as many real ones as possible. Which one matters more depends on whether false alarms or missed cases hurt more.

How to answer in an interview

Gradient descent is an optimization method that updates model parameters step by step to reduce a loss function. It follows the direction of steepest decrease, using the gradient to decide how to move. The learning rate controls how large each step is; if it is too small, learning is slow, and if it is too large, training can overshoot or diverge. Variants like stochastic gradient descent and mini-batch training are commonly used because they scale better to large datasets.

Explain like I am 5

Gradient descent is like walking downhill in small steps to reach the lowest spot in a valley. The learning rate is how big each step is.

How to answer in an interview

Feature engineering is the process of turning raw data into inputs that are more useful for a model. That can include encoding categories, scaling numbers, extracting date parts, creating ratios or interactions, and cleaning noisy values. Good feature engineering often improves model quality more than simply choosing a more complex algorithm, because the model can only learn from the signal you give it. Domain knowledge matters because it helps identify which transformations reflect the real problem.

Explain like I am 5

Feature engineering is like chopping and preparing ingredients before cooking. The model understands the meal better when the ingredients are ready in the right shape.

How to answer in an interview

A random forest builds many decision trees on random subsets of the data and random subsets of features, then combines their predictions through averaging or majority voting. This approach reduces overfitting compared to a single tree by introducing diversity among trees. The randomness in both data sampling and feature selection prevents the ensemble from relying too heavily on any one feature or pattern. Random forests are robust, require relatively little tuning, handle high-dimensional data well, and provide feature importance scores that aid interpretability.

Explain like I am 5

A random forest is like asking many different people to guess, each looking at different clues, and then taking the group vote. The group is usually more accurate than any single person.

How to answer in an interview

Gradient boosting builds trees sequentially, where each new tree focuses on correcting the errors made by the previous trees. It minimizes a loss function by adding weak learners that predict the negative gradient of the error. Unlike random forests, which build trees independently and reduce variance, boosting reduces bias by iteratively improving on hard examples. Popular implementations include XGBoost, LightGBM, and CatBoost. Gradient boosting often achieves higher accuracy than random forests but requires more careful tuning and is more prone to overfitting without regularization.

Explain like I am 5

Gradient boosting is like a team where each new member focuses on the mistakes the previous members made. Over time, the team gets better at the hard cases.

How to answer in an interview

Regularization adds constraints or penalties to a model to reduce overfitting and improve generalization. L1 regularization adds a penalty proportional to the absolute value of weights, which encourages sparsity and can perform feature selection. L2 regularization adds a penalty proportional to the squared weights, which shrinks coefficients but keeps them non-zero. Dropout randomly deactivates neurons during training, forcing the network to learn redundant representations. Early stopping is another form of regularization that halts training before the model memorizes the training data. The right choice depends on the model architecture and the nature of the data.

Explain like I am 5

Regularization is like telling the model not to rely too heavily on any one clue. It keeps the model simpler and less likely to memorize the training data.

How to answer in an interview

Imbalanced datasets occur when one class is much rarer than others, which is common in fraud detection, medical diagnosis, and anomaly detection. I handle this through resampling techniques like oversampling the minority class, undersampling the majority class, or synthetic data generation with SMOTE. I also use appropriate metrics like precision, recall, F1, or AUC-PR instead of accuracy. Cost-sensitive learning assigns higher penalties to misclassifying the minority class. In extreme cases, I use anomaly detection approaches that do not assume balanced class distributions.

Explain like I am 5

When one type of answer is very rare, the model might ignore it. I balance the data by giving more examples of the rare type or adjusting the model to pay more attention to it.

How to answer in an interview

Batch normalization normalizes the activations of a layer within each training mini-batch, typically by centering them around zero and scaling to unit variance. It reduces internal covariate shift, which stabilizes training and allows higher learning rates. Batch normalization acts as a mild regularizer, smooths the loss landscape, and often accelerates convergence. During inference, it uses running averages learned during training. I use it in deep networks where training is unstable or slow, though layer normalization is sometimes preferred in recurrent or transformer architectures.

Explain like I am 5

Batch normalization is like normalizing the grades in each class so that one hard test does not throw off the whole system. It helps the model learn more smoothly.

How to answer in an interview

L1 regularization adds the sum of absolute weights to the loss function, which can drive some weights exactly to zero, effectively performing feature selection. L2 regularization adds the sum of squared weights, which shrinks weights toward zero but rarely makes them exactly zero. L1 produces sparse models and is useful when many features are irrelevant. L2 produces smoother models and handles correlated features better. Elastic net combines both. In practice, I use L1 when I want a simpler, interpretable model and L2 when I want to prevent any single feature from dominating.

Explain like I am 5

L1 is like a strict budget that can force you to drop some features entirely. L2 is like a gentler push that keeps all features but makes sure none become too powerful.

How to answer in an interview

A support vector machine finds the optimal hyperplane that maximizes the margin between classes in a high-dimensional feature space. It is particularly effective for high-dimensional data and cases where the number of features exceeds the number of samples. The kernel trick allows SVMs to model non-linear decision boundaries by implicitly mapping data into higher dimensions without computing the transformation explicitly. SVMs are strong when margins are clear and data is not extremely noisy. They are less practical for very large datasets due to quadratic training complexity, but they remain valuable for text classification, bioinformatics, and other high-dimensional problems.

Explain like I am 5

An SVM draws the widest possible line or border between different groups. It works well when the groups are clearly separated in a high-dimensional space.

How to answer in an interview

Dimensionality reduction transforms high-dimensional data into a lower-dimensional representation while preserving as much relevant structure as possible. PCA finds orthogonal directions of maximum variance and projects data onto them. t-SNE and UMAP are non-linear methods that preserve local neighborhood structure, useful for visualization. Feature selection methods like mutual information or L1-based selection remove irrelevant features entirely. Reducing dimensions improves computational efficiency, reduces overfitting, and helps visualization. The choice between linear and non-linear methods depends on the complexity of the data structure.

Explain like I am 5

Dimensionality reduction is like summarizing a long book into key points. You lose some details, but the main ideas are preserved and easier to work with.

How to answer in an interview

Bias is error from making assumptions that are too simple, which can cause underfitting. Variance is error from a model changing too much based on the training data, which can cause overfitting. The tradeoff is that more flexible models usually reduce bias but increase variance, while simpler models do the opposite. In practice, I manage this tradeoff with model choice, regularization, feature engineering, cross-validation, and enough high-quality data.

Explain like I am 5

Bias is when a model is too simple and keeps missing the point. Variance is when it changes its mind too much from one example to the next.

How to answer in an interview

Deploying a machine learning model involves packaging the model into a serving infrastructure that can handle real-time or batch predictions reliably. I typically containerize the model, expose it through a REST or gRPC endpoint, and set up load balancing and auto-scaling. I establish CI/CD pipelines that retrain and validate models automatically. Monitoring is critical: I track prediction distributions, latency, error rates, and data drift. I maintain rollback capability and shadow deployments for safe rollouts. The key is treating the model as a software artifact with versioning, testing, and operational rigor.

Explain like I am 5

Deploying a model means putting it into an app so real users can benefit from it. I need to make sure it is fast, reliable, and I can update or roll it back safely.

How to answer in an interview

A/B testing for ML models involves randomly assigning users to either the current model or a new candidate model and comparing their performance on real-world metrics. I define a hypothesis, choose a primary metric, calculate the required sample size for statistical power, and run the experiment for a sufficient duration to capture weekly patterns. I analyze results using statistical tests to determine significance, check for novelty effects and sample ratio mismatch, and examine segment-level effects. The goal is to establish a causal link between the model change and the business outcome, not just correlation.

Explain like I am 5

An A/B test is like a fair experiment where half the users get the new model and half get the old one. We compare the results to see if the new model is actually better.