Question 1

What are the best Python libraries for machine learning?

Accepted Answer

The essential Python ML stack is: NumPy (array math), Pandas (data manipulation), scikit-learn (classical ML: classification, regression, clustering, preprocessing, pipelines, evaluation), Matplotlib/Seaborn (visualization), TensorFlow/Keras (deep learning, production-grade), PyTorch (research + HuggingFace models), HuggingFace Transformers (pretrained NLP models), LangChain (LLM apps and RAG pipelines), and XGBoost/LightGBM (gradient boosting for tabular data). Start with NumPy + Pandas + scikit-learn. Add TensorFlow or PyTorch when you need neural networks. Add LangChain when you work with LLMs.

Question 2

How do you handle imbalanced datasets in machine learning?

Accepted Answer

Imbalanced datasets (e.g., 95% not-fraud, 5% fraud) require several strategies: (1) Choose the right metric — accuracy is misleading; use precision, recall, F1, or ROC-AUC instead. (2) Resampling — oversample the minority class with SMOTE (Synthetic Minority Oversampling Technique) or undersample the majority class. (3) Class weights — set class_weight="balanced" in scikit-learn classifiers, which automatically adjusts loss penalties. (4) Threshold tuning — instead of defaulting to 0.5 probability threshold, tune the classification threshold to trade off precision vs recall for your business objective. (5) Ensemble methods like Random Forest and XGBoost handle imbalance better than Logistic Regression out of the box.

Question 3

Should I use a CNN or RNN for sequence data?

Accepted Answer

For most sequence tasks in 2025, use Transformers (not CNNs or RNNs). But the traditional guidance: CNNs are preferred when local patterns matter (e.g., 1D CNNs for short time series, n-gram detection in text) because they process sequences in parallel with convolutional filters. RNNs (specifically LSTMs and GRUs) are preferred for long-range temporal dependencies where order matters critically (e.g., predicting future values in a multi-step time series). In practice, 1D CNNs often outperform RNNs on time-series classification while training faster. For variable-length text sequences, Transformers (BERT, GPT) dominate because their self-attention mechanism captures arbitrary-distance dependencies without the vanishing gradient problem that affects RNNs.

Question 4

When should I use RAG versus fine-tuning for LLM applications?

Accepted Answer

Use RAG (Retrieval-Augmented Generation) when: your data changes frequently (product catalogs, policy documents, research papers), you have large private knowledge bases, or you need the model to cite sources. RAG requires no model retraining — just index your documents and retrieve relevant chunks at query time using vector similarity. Use fine-tuning when: you need to deeply change the model's tone, format, or domain expertise, when you have high-quality instruction datasets (input-output pairs), or when you need consistent behavior that prompting alone cannot achieve. LoRA and QLoRA (quantized LoRA) from HuggingFace PEFT make fine-tuning feasible on consumer GPUs by training only 0.1–1% of model parameters. The decision tree: Start with prompting → add RAG for private knowledge → fine-tune only if RAG + prompting still falls short.

Question 5

What is the difference between an ML engineer and a data scientist career path?

Accepted Answer

ML Engineers and Data Scientists both build models but focus on different parts of the lifecycle. Data Scientists focus on analysis, experimentation, and insight: they clean data, explore patterns, build models in Jupyter notebooks, and communicate findings to stakeholders. They need strong statistics, SQL, visualization, and domain knowledge. ML Engineers focus on building production ML systems: they take models from notebooks and deploy them as reliable, scalable services. They need strong software engineering (Python, APIs, Docker, cloud), model optimization, and MLOps skills (pipelines, CI/CD, monitoring). AI Engineers is an emerging role focused on building LLM-powered products using APIs, RAG, and agents — closer to backend engineering than traditional ML. MLOps Engineers build the infrastructure layer: model registries, feature stores, training pipelines, and monitoring systems. Salary ranges are similar; ML Engineers and AI Engineers tend to earn slightly more due to engineering scope.