Machine learning interview questions for campus hiring — models, metrics, and real projects.
Traditional: explicit rules coded by developers. ML: model learns patterns from data to make predictions. Example: spam filter learns from labeled emails instead of hand-written keyword lists. ML powers recommendations, fraud detection, and computer vision in modern IT products.
Supervised: labeled data (input → known output) — classification, regression. Unsupervised: find structure without labels — clustering, dimensionality reduction. Reinforcement: agent learns via rewards in an environment — robotics, game AI, some ad ranking.
Model memorizes training data, poor on new data. Signs: high training accuracy, low validation accuracy. Fixes: more data, regularization (L1/L2), simpler model, dropout (neural nets), cross-validation, early stopping. Very common ML interview question for campus roles.
Train: fit model parameters. Validation: tune hyperparameters and compare models without peeking at test. Test: final unbiased estimate once before launch. Random split or stratified for imbalanced classes; time-based split for forecasting.
Classification predicts discrete labels (pass/fail, spam/ham). Regression predicts continuous values (house price, temperature). Metrics differ: accuracy/F1 vs MAE/RMSE. Many business problems map clearly to one type.
Precision: of predicted positives, how many are correct — important when false positives costly. Recall: of actual positives, how many caught — important when missing positives costly (fraud, disease). F1: harmonic mean balancing both. Use case drives which metric to optimize.
Table of TP, FP, TN, FN for binary classification. Visualizes error types beyond single accuracy — critical when classes imbalanced (99% non-fraud). Interviewers ask you to interpret each quadrant.
Split data into k folds; train on k-1, validate on 1, rotate k times. Average performance reduces variance of single split estimate. k=5 or 10 common on tabular datasets in sklearn workflows.
Linear regression predicts continuous output with linear combination of features. Logistic regression predicts probability of class using sigmoid — still 'linear' in feature space but used for binary classification. Baseline models interviewers expect you to explain.
Optimization algorithm minimizing loss by updating weights in direction of negative gradient. Learning rate controls step size — too high diverges, too low slow. Variants: SGD, mini-batch, Adam. Foundation for training neural networks.
Layers of neurons applying weighted sums + nonlinear activations (ReLU). Deep nets learn hierarchical features. Need large data and compute; prone to overfitting without regularization.
Transfer learning (pretrained models) dominates vision/NLP before LLM era and still for specialized tasks.
Creating informative inputs: encoding categories, scaling numerics, date parts, text TF-IDF, domain ratios. Good features beat complex models on small tabular data — common in analytics-heavy service company case studies.
High bias: underfitting, too simple. High variance: overfitting, too complex. Goal: minimize total error on unseen data. Regularization and model selection manage this tradeoff.
Training uses information unavailable at prediction time (future data, target-encoded without proper CV, test set in preprocessing). Inflates metrics and fails in production. Interview gold: mention pipeline fit only on train fold.
Serialize model (pickle/ONNX), wrap in REST API, containerize, monitor drift and latency, retrain on schedule or trigger. A/B test new model vs old. Tools: MLflow, SageMaker, BentoML — awareness suffices for many fresher loops.
AI: broad goal of intelligent systems. ML: subset learning from data. Deep learning: ML using deep neural networks. LLMs are deep learning; not all AI is generative LLMs — clarify in interviews.
Python stack: NumPy, Pandas, scikit-learn, Matplotlib/Seaborn; optionally PyTorch or TensorFlow for deep learning. Mention one project: dataset → EDA → model → metric → improvement idea.
Techniques: class weights, SMOTE oversampling, threshold tuning, use PR-AUC instead of accuracy, collect more minority samples. Real fraud and medical datasets skew heavily — Indian service ML teams ask this often.
Collaborative filtering: users similar to you liked X. Content-based: items similar to what you viewed. Hybrid combines both. Netflix/Amazon/Flipkart use large-scale variants — good system design + ML crossover answer.
LLMs are pre-trained on text then adapted via prompts/fine-tuning; classical ML often trains smaller models on tabular or curated datasets. Products combine both: LLM for language, ML for ranking/fraud, rules for compliance.
