Machine Learning Interview Questions for India 2026

Core concepts tested at every level: Supervised vs unsupervised vs reinforcement learning. Bias-variance tradeoff: high bias = underfitting, high variance = overfitting. Cross-validation: k-fold to estimate generalisation error. Regularisation: L1 (Lasso: sparse weights, feature selection) vs L2 (Ridge: shrinks all weights). Gradient descent: batch, stochastic, mini-batch. Learning rate schedules. Overfitting detection and remedies: regularisation, dropout (neural nets), more data, early stopping. Feature scaling: StandardScaler for gradient-based models, MinMaxScaler for distance-based models (KNN, SVM): tree-based models do not need scaling.

Algorithms tested at Indian ML interviews: Linear and logistic regression (assumptions, interpretability, when to use). Decision trees (information gain, Gini impurity, pruning). Random forest (ensemble of trees, out-of-bag error, feature importance). Gradient boosting: XGBoost, LightGBM, CatBoost: the go-to for structured/tabular data at Indian product companies. SVM: kernel trick, support vectors, C and gamma parameters. KNN: lazy learner, curse of dimensionality. Neural networks: architecture (layers, activations), backpropagation, common architectures (CNN for images, LSTM/Transformer for sequences). For recommendation systems (Flipkart, Zomato, Swiggy use cases): collaborative filtering, matrix factorisation, content-based filtering.

Model evaluation metrics are heavily tested: interviewers expect you to know which metric to use when and why: Accuracy: misleading for imbalanced datasets. Precision: out of predicted positives, how many are correct? Recall (Sensitivity): out of actual positives, how many did we catch? F1 score: harmonic mean of precision and recall: use when both matter. ROC-AUC: rank-ordering ability, threshold-independent: use to compare models. PR-AUC: better than ROC-AUC for highly imbalanced datasets (fraud detection, rare disease prediction). RMSE, MAE for regression: MAE is more robust to outliers. Business metric alignment: always connect model metrics to business outcomes: reducing churn prediction false negatives reduces revenue loss.

MLOps questions appear at senior ML roles at Indian product companies: Feature stores: centralised feature management (Feast, Tecton): prevents training-serving skew. Training-serving skew: features computed differently during training vs inference: the most common production ML bug. Model versioning: MLflow, Weights and Biases. CI/CD for ML: automated retraining pipelines, data validation (Great Expectations), model performance monitoring (data drift, concept drift). A/B testing ML models: holdout groups, traffic splitting, statistical significance. Model serving: REST API (Flask/FastAPI), model as a microservice, ONNX for cross-framework deployment, TorchServe, TensorFlow Serving. Shadow deployment: running a new model in parallel to compare against production before switching.