Extreme multi-class classification (tens of thousands of classes)

[MaxHalford]

• Extreme F-Measure Maximization using Sparse Probability Estimates
• Probabilistic Label Trees for Extreme Multi-label Classification

Filter trees

Filter trees are a way to reduce multi-class classification to binary classification (like one-vs-all). Intuitively, the classes are recursively into two parts, therefore defining a tree, which only requires log(n) steps for predicting and training. Multiple filter trees become a so-called error-correcting tournament. See this page from Vowpal Wabbit.