Classical ML & Statistics

Classical Machine Learning encompasses the foundational algorithms and statistical techniques that remain essential in production ML systems. Despite the rise of deep learning, classical ML methods are often preferred for tabular data, interpretability requirements, and resource-constrained environments.

Key areas include supervised learning (linear/logistic regression, decision trees, random forests, gradient boosting with XGBoost/LightGBM), model evaluation (cross-validation, bias-variance tradeoff, precision-recall, ROC-AUC), and feature engineering (feature selection, encoding, scaling, handling missing data).

Understanding these fundamentals is critical: overfitting vs. underfitting, regularization (L1/L2), ensemble methods, and interpretability tools (SHAP, feature importance). Many interview questions test these concepts because they reveal deep understanding of how and why ML algorithms work.

Overview

ML Concepts

What is overfitting? How do you prevent it?

Explain the bias-variance tradeoff.

Explain precision, recall, and F1 score. When to optimize for which?

How do you ensure LLMs output valid structured data (JSON, specific formats)?

Explain PCA (Principal Component Analysis). When and how do you use it?

How do you choose the number of epochs? Explain early stopping and gradient descent optimizers.

Deep-Dive Concepts (from Projects)

Matrix Factorization Deep Dive

Two-Stage Recommendation Architecture

Recommendation Evaluation Metrics

Cold-Start Problem Solutions

Diversity in Recommendations

Implicit Feedback in Recommendations

Contrastive Learning (MoCo)

Self-Distillation (DINO)

Grad-CAM for Multi-Label Classification

Domain Shift in Medical Imaging

Test-Time Adaptation (TENT)

Confidence Calibration

Feature Engineering Best Practices

Target Encoding Deep Dive

XGBoost for Tabular Classification

SHAP Values and Model Explainability

Palantir Foundry Ontology Architecture

Star Schema Design for Analytics

Cost-Sensitive Classification

Feature Engineering for Behavioral Data