8 - Supervised Learning - Discriminative Classification英文文献资料.pdfVIP

Machine Learning, Machine Learning (extended) 8 – Supervised Learning: Discriminative Classification Kashif Rajpoot k.m.rajpoot@cs.bham.ac.uk School of Computer Science University of Birmingham Outline • Supervised learning • Classification • Discriminative classification • Decision boundary • The margin • Maximizing the margin • Making predictions • Support vectors • Hard margin • Soft margin • Non-linear decision boundary 2 • Kernel trick Supervised learning • Regression • Minimised loss (e.g. least squares) • Maximum likelihood • Classification • Generative (e.g. Bayesian) • Instance-based (e.g. k-NN) • Discriminative (e.g. SVM) 3 Classification • A set of N objects with attributes (usually vector) • Each object has an associated target label • Binary classification ∈ 0, 1 or ∈ −1, 1 • Multi-class classification ∈ 1,2, …, • Classifier learns from , , …, and , , …, so 1 2 1 2 that it can later classify 4 Generative vs discriminative classification • Generative classifiers generate a model for each class, based on training samples available • Data in each class can be seen as generated by some model • For new test samples, they assign these samples to the class that suits best (e.g. by probability measure) • In contrast, discriminative classifiers attempt to explicitly define the decision boundary that separates the classes • Intuitively, these methods are for binary class problems but can be extended to multi-class problems