Learning Coordination Classifiers Guo, Greiner, and Schuurmans University of Alberta Presented by Nick Rizzolo
Outline Standard assumptions about classification What’s a “coordination classifier”? How is it trained? How is it evaluated? How is this approach justified? Experiments 9/13/05 AIML Seminar
Standard Classification Assumptions Input: Output: Training and testing data are independent and identically distributed (i.i.d.) f 9/13/05 AIML Seminar
Coordination Classification Pair examples Multi-class f Coordination 9/13/05 AIML Seminar
Training Amount of training examples is squared Maximum likelihood, logistic regression, naïve Bayes, Bayes networks, neural networks, etc. Trained classifier makes dependent associations 9/13/05 AIML Seminar
Evaluation HMM, CRF, voting, … What the authors did: Markov Random Field Training examples Testing examples 9/13/05 AIML Seminar
Evaluation (cont’) Full network is impractical Just train-test edges Analogous to kernel based algorithms Easy to compute most likely labeling Just test-test edges Analogous to ensemble (voting) methods Probabilistic inference is expensive Random edge subsampling 9/13/05 AIML Seminar
Rationale x1 x2 y1 y2 test examples true conditional model learned 9/13/05 AIML Seminar
Experiment 1 Logistic regression, belief propagation, only test edges, 18 edges per example 9/13/05 AIML Seminar
Naïve Bayes, belief propagation, only test edges, 18 edges per example Experiment 2 Naïve Bayes, belief propagation, only test edges, 18 edges per example 9/13/05 AIML Seminar
Experiment 3 Logistic regression, belief propagation, 18 edges per example Only train-test edges Test-test and train-test edges 9/13/05 AIML Seminar
Neural Network, voting, only test edges, 18 edges per example Experiment 4 Neural Network, voting, only test edges, 18 edges per example 9/13/05 AIML Seminar
Logistic regression, belief propagation, only test edges Experiment 5 Logistic regression, belief propagation, only test edges 9/13/05 AIML Seminar