110 likes | 123 Views
This module explores various techniques such as SVM, neural networks, and genetic algorithms for filtering and recommendation systems using supervised machine learning. Learn about linear separators, document vectors, and how to balance adaptation with generalization in training classifiers. Discover strategies to prevent overtraining and optimize performance on new data. Understand the challenges of filtering in dynamic content and leverage behavioral signals for decision-making. Master threshold learning and explore innovative approaches for effective filtering.
E N D
Filtering and Recommendation INST 734 Module 9 Doug Oard
Agenda • Filtering • Recommender systems • Classification
Supervised Machine Learning • Model-based techniques • Hill climbing (e.g., Rocchio) • Statistical classification (e.g., SVM) • Rule induction (e.g., decision trees) • Neural networks (e.g., “deep learning”) • Population-Based Techniques • Instance-based learning (e.g., kNN) • Genetic algorithms
Adaptive Vector-Space Filtering New Documents Make Document Vectors Compute Similarity Vectors Ranking Threshold Select and Examine (user) Document, Vectors Vector(s) Assign Ratings (user) Rating, Vector Initial Profile Features Make Profile Vector Update User Model Vector
Latent Semantic Indexing New Documents Sparse Vectors Dense Vectors Make Document Vectors Reduce Dimensions Compute Similarity Ranking Threshold Matrix Representative Documents Sparse Vectors Make Document Vectors SVD Select and Examine (user) Document, Dense Vector Dense Vector(s) Assign Ratings (user) Matrix Rating, Dense Vector Initial Profile Features Sparse Vector Dense Vector Make Profile Vector Reduce Dimensions Update User Model
Linear Separators • Which of the linear separators is optimal? Original from Ray Mooney
Maximum Margin ClassificationSupport Vector Machine (SVM) • Implies that only support vectors matter • Other training examples are ignorable. Original from Ray Mooney
Soft Margin SVM ξi ξi Original from Ray Mooney
Non-linear SVMs Φ: x→φ(x) Original from Ray Mooney
Training Supervised Classifiers • All learning systems share two problems • They need some “inductive bias” • They must balance adaptation with generalization • Overtraining can hurt performance • Performance on training data rises and plateaus • Performance on new data rises, then falls • Useful strategies • Hold out a “devtest” set to find peak on new data • Emphasize exploration early, exploration later
Summing Up • Filtering poses some unique challenges • Adversarial behavior, new terms, throughput • Behavioral signals offer unique opportunities • For both static and dynamic content • Supervised classifiers learn to make decisions • Two-sided training • Threshold learning