Model answers for 4 ML questions
Typical round: fundamentals, a case study, a project deep dive
Q: Explain the bias-variance trade-off. A: Bias is error from a model too simple to capture the pattern; variance is error from a model too sensitive to the training data. Increasing complexity lowers bias but raises variance. You find the balance with validation curves, regularisation and more data.
Q: Your fraud model has 99% accuracy. Is it good? A: Not necessarily. If only 1% of transactions are fraud, predicting 'not fraud' always gives 99%. Use precision, recall, F1 or PR-AUC, and pick a threshold based on the cost of missed fraud vs false alarms.
Q: How do you prevent overfitting? A: More data, simpler models, regularisation such as L1 or L2, dropout for neural nets, early stopping, cross-validation, and removing leaky or noisy features.
Q: Why might a model perform well offline but badly in production? A: Data drift, training-serving skew in features, data leakage during training, or a validation split that didn't match real-world time order. Monitoring input distributions and live metrics catches this.