Data ScientistInterview Answers

I start with problem framing — what decision will the model inform, and is ML actually the right tool? I explore the data, understand distributions, handle missingness, and engineer features informed by domain knowledge. I select a baseline model first (always start simple), then experiment with more complex approaches. I validate rigorously — using cross-validation, holding out a test set, and evaluating with appropriate metrics for the problem (not just accuracy). Finally, I consider deployment: how will predictions be consumed, and how will model performance be monitored over time?

First, I choose appropriate evaluation metrics — accuracy is useless with imbalanced classes. I use precision, recall, F1, AUC-ROC, or Precision-Recall AUC depending on the cost of false positives vs false negatives. On the data side, I use SMOTE for oversampling, undersampling of the majority class, or class weighting in the model. I also experiment with threshold adjustment — the default 0.5 probability threshold isn't always optimal. I evaluate which approach works best through cross-validated experiments.

Bias is error from overly simplistic assumptions — underfitting. A high-bias model gets the same wrong answer regardless of training data. Variance is error from overfitting to noise in training data — it performs well on training data but poorly on new data. The tradeoff is that reducing bias (more complex models) increases variance and vice versa. Regularisation (L1/L2), cross-validation, and ensemble methods are tools for managing this tradeoff. In practice, I start simple, measure training vs validation error, and add complexity only when there's evidence of underfitting.

Feature engineering is often more impactful than model selection. I start with domain knowledge — what variables would a subject matter expert say are predictive? I then explore the data: distributions, correlations, and potential interaction effects. I create features like lagged variables for time series, interaction terms, polynomial features for non-linear relationships, and encoded representations for categorical variables. I validate each feature's contribution using feature importance plots and ablation studies — features that don't add signal should be removed to reduce overfitting.

I translate model outputs into business impact. Instead of "the model has an AUC of 0.82," I say "if we use this model to prioritise our outreach, we'd identify 80% of churning customers while only contacting 30% of our base — saving the team X hours and recovering an estimated Y in revenue." I use visualisations — confusion matrices in plain English, lift charts, SHAP plots for explainability. I always include confidence and limitations. Decision-makers need to trust the model, not just use it.

I've applied deep learning for NLP (BERT, fine-tuning transformers for text classification), computer vision (CNNs for image classification), and time series forecasting (LSTMs, temporal convolutional networks). I use PyTorch and TensorFlow/Keras. I'm also familiar with transfer learning and know when not to use deep learning — when data is limited, simpler models often outperform neural networks. I keep up with the field through papers (ArXiv), fast.ai, and implementing new architectures on personal projects.

I start by understanding the protected attributes in the context and checking for training data bias — historical data often encodes historical discrimination. I use fairness metrics (equalized odds, demographic parity, individual fairness) appropriate to the use case and accept that different fairness criteria can be mathematically in tension. I conduct disparate impact analysis across demographic groups and document findings. I also involve ethicists and domain experts in high-stakes applications. Bias in models is a business and ethical risk, not just a technical problem.

I've deployed models as REST APIs (FastAPI, Flask) containerised with Docker and orchestrated with Kubernetes. I use MLflow for experiment tracking and model registry. I've built monitoring pipelines to detect data drift (using Evidently, Great Expectations) and model performance degradation over time. I write unit tests for data pipelines and model inference functions, and treat model code with the same engineering standards as application code. Deployment is where most ML projects succeed or fail.