Machine Learning · Beginner interview

Machine learning is a way to get computers to learn patterns from data rather than following explicitly coded rules. In traditional programming you write the rules and the program applies them; in ML you give examples (data and outcomes) and an algorithm infers the rules (a model) that map inputs to outputs. You use it for problems too complex or fuzzy to hand-code, like image recognition or spam detection.

Features are the input variables the model learns from — the measurable properties of each example (e.g. a house's size, location, age). The label is the target you want to predict (e.g. the house price). In supervised learning the model learns the relationship between features and labels; choosing and engineering good features is often more important than the algorithm.

The training set is used to fit the model. The validation set is used during development to tune hyperparameters and compare models without touching the test data. The test set is held out until the end to estimate how the final model performs on unseen data. Keeping them separate prevents you from fooling yourself — evaluating on data the model has seen overstates performance.

Supervised learning trains on labeled examples to predict outputs (classification, regression). Unsupervised learning finds structure in unlabeled data (clustering, dimensionality reduction). Reinforcement learning trains an agent to take actions in an environment to maximize a reward signal through trial and error. The difference is what signal guides learning: labels, structure, or rewards.

Both are supervised, but classification predicts a discrete category (spam/not-spam, which digit) while regression predicts a continuous number (price, temperature). They use different loss functions and metrics — accuracy/F1 for classification, error metrics like RMSE for regression. The nature of the target variable decides which you're doing.

Linear regression (predict a number) and logistic regression (binary classification) are the foundational linear models. Decision trees split data on feature thresholds and are interpretable. k-Nearest Neighbors classifies by the closest examples. k-Means clusters unlabeled data. Knowing when each applies — and that simpler models are often a strong baseline — matters more than memorizing many algorithms.

Overfitting is when a model learns the training data too well — including its noise — so it performs great on training data but poorly on new data (high variance). Underfitting is when a model is too simple to capture the pattern, doing poorly even on training data (high bias). You detect overfitting by a large gap between train and test performance, and address it with more data, simpler models, or regularization.

To estimate how the model generalizes to data it hasn't seen. If you evaluate on the same data you trained on, the model can score well by memorizing rather than learning the pattern, giving a falsely optimistic result. Holding out a test set simulates real deployment and reveals overfitting.

On imbalanced datasets. If 99% of emails are not spam, a model that always predicts 'not spam' is 99% accurate but useless — it never catches spam. In those cases you look at precision (of predicted positives, how many were right) and recall (of actual positives, how many you caught), and combine them with F1. Accuracy only tells the full story when classes are balanced and errors are equally costly.

Feature scaling puts features on a comparable range — normalization (0–1) or standardization (zero mean, unit variance). It matters for algorithms that use distances or gradients (kNN, SVM, neural nets, gradient descent), where a feature with a large range would otherwise dominate. Tree-based models don't need it. You fit the scaler on training data and apply it to validation/test to avoid leakage.

Cross-validation evaluates a model by splitting the data into k folds, training on k−1 and validating on the held-out fold, rotating through all folds and averaging the results. It gives a more reliable performance estimate than a single split — especially on small datasets — and reduces the chance that a lucky/unlucky split misleads you. It's commonly used for model selection and hyperparameter tuning.

Define the problem and metric; collect and clean data; explore and engineer features; split into train/validation/test; choose and train a model; evaluate and tune; then deploy and monitor. It's iterative — results send you back to features or data. Most of the effort is in data preparation and evaluation, not the modeling step itself.