Machine Learning: A Prerequisite Map
Start with how data is evaluated and what a model optimizes. Then choose a branch: regression, trees or learned representations. Ensembles build on trees; retrieval builds on representations. You can study independent branches in either order.
The explorer follows this prerequisite order. The map below makes the connections explicit. Each arrow recommends an introductory prerequisite, not every chapter in that folder. Advanced methods and paper companions are there when your next question needs them.
Choose a destination
Select a topic to highlight the foundations leading to it. The route lists those topics in reading order; you can skip any you already know. Other branches stay available.
Your reading route
Three useful starting routes
A first tabular model
- Choose the split and understand what counts as unseen data.
- Read data-preparation foundations and loss versus metric.
- Start with linear regression. For a binary target, continue to logistic regression, using cross-entropy when the loss needs context.
- Use cross-validation to compare choices and pipelines to keep fitted preprocessing inside the training boundary.
Trees and risk models
- After data and loss foundations, study decision trees.
- Continue to bagging and boosting (functional gradient descent, AdaBoost, Gradient Boosting, and XGBoost). Review regression and loss gradients before the boosting derivations.
- Branch into anomaly detection after trees, or graph risk modeling after a tabular baseline.
- For explanations, use the explainability guide. Shapley values come before TreeSHAP; linear fitting helps with LIME.
Embeddings, search and evaluation
- Review embedding-table lookup and entropy and log-loss.
- Learn behavioral embeddings, including how to evaluate their neighborhoods.
- Study approximate nearest-neighbor search and its benchmarks.
- Continue to search and retrieval evaluation: relevance labels, ranking metrics, uncertainty, online experiments and RAG evaluation.
Work through the handwritten notes
The Handwritten Notes collection brings together eight expanded guides: MLE and MAP, linear regression, Naïve Bayes, logistic regression, K-means, PCA, LDA projection, and LDA classification. Each uses visible data tables and complete calculations, with the original PDFs and selected page scans alongside the explanation.
Keep the first pass small
The data-preparation folder is a toolbox. Begin with splitting, cleaning, missing values and simple representations. Return to feature selection after you have a baseline to compare, and to imbalance after classification metrics and thresholds. Use dimensionality reduction when you have a specific compression, prediction or visualization goal.
Likewise, transport and triplet losses are optional branches of the loss collection. The handwritten tree notes provide a parallel companion. Recommended papers extend topics you have already met; they are not a gate before fitting a first model.
If a mathematical prerequisite is unfamiliar, revisit probability foundations or numerical optimization when the relevant chapter calls for it.