Machine learning
Core algorithms, honest evaluation and real-world applications of machine learning, in one map.
Overview
Machine learning (ML) is the practice of building systems that learn patterns from data instead of following hand-written rules. Given examples, such as past transactions labelled as fraud or not, an ML model learns to make predictions on new cases.
A new employee learns to spot suspicious invoices by reviewing hundreds of past cases with a mentor; machine learning does the same with data, at scale and consistently. This page maps the foundations, the methods that work best on business data, how to evaluate honestly, and where ML creates value across industries.
Foundations
Every practitioner should understand a handful of classic methods: Linear regression and Logistic regression as interpretable baselines, Decision trees and Gradient boosting for tabular data, k-nearest neighbors and Naive Bayes as simple, fast learners, and Principal component analysis and t-SNE for understanding high-dimensional data. Broader families are covered in Supervised learning, Unsupervised learning and Reinforcement learning.
How models learn
Almost every model is trained the same way: pick a loss function that scores mistakes, then lower it step by step with gradient descent. Good results depend as much on the data as on the algorithm: data preprocessing and feature engineering shape what the model sees, regularization keeps it from memorising, the bias-variance trade-off explains under- and overfitting, hyperparameter tuning finds good settings, and evaluation metrics say whether it actually works.
Classic models
SVM, Random forest, Clustering and Isolation forest remain everyday tools. Ensemble learning explains why combining many models, as boosting libraries do, so often wins.
Learning paradigms
How much labelled data you have decides the approach: Supervised, Semi-supervised, Self-supervised and Unsupervised learning, learning from rewards in Reinforcement learning, and reusing knowledge with Transfer learning, Few-shot and Zero-shot learning.
Tabular data and foundation models
Most business data lives in tables. Boosted trees such as XGBoost, LightGBM and CatBoost have long dominated; tabular foundation models such as TabPFN and the large tabular foundation model documented in this site's docs now predict on new tables in context, without training a model per table. The detailed guide is Tabular foundation models, with In-context learning explaining the mechanism.
Evaluating honestly
Good models are proven, not assumed. Cross-validation estimates generalisation; held-out reports, leakage checks and calibrated metrics keep results honest; explainability tools such as SHAP and LIME show why a model decides as it does. See also Evaluation in the model docs.
Where ML is applied
Industrial and business applications range from Predictive maintenance and Digital twins to credit and fraud in FinTech, network operations in Telecom, forecasting in Logistics and personalisation in E-commerce.
The full map
ML foundations
Learning paradigms
Training fundamentals
Classic models
Industries and applications
Related topics
Frequently asked questions
What is the difference between AI and machine learning?
AI is the broad goal of building intelligent systems; machine learning is the main approach used today, in which systems learn from data.
Which ML method works best on tables?
Gradient-boosted trees are a strong default; tabular foundation models are increasingly competitive and need no per-table training.
Summary
Machine learning turns data into predictions. Strong foundations, methods suited to tabular data, honest evaluation and clear business problems are what make it work in practice.