COMP 328: Final Review Spring 2010 Nevin L. Zhang Department of Computer Science & Engineering The Hong Kong University of Science & Technology http://www.cse.ust.hk/~lzhang/ Can be used as cheat sheet Page 2 Pre-Midterm Algorithms for supervised learning Decision trees Instance-based learning Naïve Bayes classifiers Neural networks Support vector machines General issues regarding supervised learning Classification error and confidence interval Bias-Variance tradeoff PAC learning theory Post-Midterm Clustering Distance-Based Clustering Model-Based Clustering Dimension Reduction Principal Component Analysis Reinforcement Learning Ensemble Learning Clustering Distance/Similarity Measures Distance-Based Clustering Partitional and Hierarchical clustering K-Means: Partitional Clustering K-Means: Partitional Clustering Different initial points might lead to different partitions Solution: Multiple runs, Use evaluation criteria such as SSE to pick the best one Hierarchical Clustering Agglomerative and Divisive Cluster Similarity Cluster Validation External indices Entropy: Average purity of clusters obtained Mutual Information between class label and cluster label Cluster Validation External Measure Jaccard Index Rand Index Measure similarity between two relationships: in-same-class & in-same-cluster # pairs in same cluster # pairs in diff cluster # pairs w/ same label a b # pairs w/ diff label c d Cluster Validation Internal Measure Dunn’s index Cluster Validation Internal Measure Post-Midterm Clustering Distance-Based Clustering Model-Based Clustering Dimension Reduction Principal Component Analysis Reinforcement Learning Ensemble Learning Model-Based Clustering Assume data generated from a mixture model with K components Estimate parameters of the model from data Assign objects to clusters based posterior probability: Soft Assignment Gaussian Mixtures Learning Gaussian Mixture Models EM EM EM l(t): Log likelihood of model after t-th iteration l(t): increases monotonically with t But might go to infinite in case of singularity Solution: place bound on eigen values of covariance matrix Local maximum Multiple restart Use likelihood to pick best model EM and K-Means K-Means is hard-assignment EM Mixture Variable for Discrete Data Latent Class Model Learning Latent Class Models Always converges Post-Midterm Clustering Distance-Based Clustering Model-Based Clustering Dimension Reduction Principal Component Analysis Reinforcement Learning Ensemble Learning Dimension Reduction Necessary because there are data sets with large numbers of attributes that are difficult to learning algorithms to handle. Principal Component Analysis PCA Solution PCA Illustration Eigenvalues and Projection Error Post-Midterm Clustering Distance-Based Clustering Model-Based Clustering Dimension Reduction Principal Component Analysis Reinforcement Learning Ensemble Learning Reinforcement Learning Markov Decision Process A model of how agent interact with its environment Markov Decision Process Value Iteration Reinforcement Learning Q-Learning Q-Learning From Q-function based value iteration Ideas In-place/asynchronous value iteration Approximate expectation using samples ε-greedy policy (for exploration/exploitation) tradeoff Time Difference Learning Sarsa is also time difference learning Post-Midterm Clustering Distance-Based Clustering Model-Based Clustering Dimension Reduction Principal Component Analysis Reinforcement Learning Ensemble Learning Ensemble Learning Bagging: Reduce Variance Boosting: Reduce Classification Error AdaBoost: Exponential Error